praeclarumjj3 commited on
Commit
6e445f1
1 Parent(s): 4786fd9

Create Space

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +4 -4
  2. README.md +6 -5
  3. configs/.DS_Store +0 -0
  4. configs/ade20k/Base-ADE20K-UnifiedSegmentation.yaml +68 -0
  5. configs/ade20k/oneformer_R50_bs16_160k.yaml +58 -0
  6. configs/ade20k/oneformer_dinat_large_IN21k_384_bs16_160k.yaml +42 -0
  7. configs/ade20k/oneformer_swin_large_IN21k_384_bs16_160k.yaml +40 -0
  8. configs/cityscapes/.DS_Store +0 -0
  9. configs/cityscapes/Base-Cityscapes-UnifiedSegmentation.yaml +68 -0
  10. configs/cityscapes/oneformer_R50_bs16_90k.yaml +59 -0
  11. configs/cityscapes/oneformer_dinat_large_bs16_90k.yaml +22 -0
  12. configs/cityscapes/oneformer_swin_large_IN21k_384_bs16_90k.yaml +20 -0
  13. configs/coco/Base-COCO-UnifiedSegmentation.yaml +54 -0
  14. configs/coco/oneformer_R50_bs16_50ep.yaml +59 -0
  15. configs/coco/oneformer_dinat_large_bs16_100ep.yaml +22 -0
  16. configs/coco/oneformer_swin_large_IN21k_384_bs16_100ep.yaml +25 -0
  17. deform_setup.sh +21 -0
  18. demo/colormap.py +170 -0
  19. demo/defaults.py +77 -0
  20. demo/predictor.py +190 -0
  21. demo/visualizer.py +1350 -0
  22. examples/ade20k.jpeg +3 -0
  23. examples/cityscapes.png +3 -0
  24. examples/coco.jpeg +3 -0
  25. gradio_app.py +194 -0
  26. oneformer/.DS_Store +0 -0
  27. oneformer/__init__.py +9 -0
  28. oneformer/config.py +239 -0
  29. oneformer/data/__init__.py +2 -0
  30. oneformer/data/bpe_simple_vocab_16e6.txt +0 -0
  31. oneformer/data/bpe_simple_vocab_16e6.txt.gz +3 -0
  32. oneformer/data/build.py +117 -0
  33. oneformer/data/dataset_mappers/__init__.py +1 -0
  34. oneformer/data/dataset_mappers/coco_unified_new_baseline_dataset_mapper.py +341 -0
  35. oneformer/data/dataset_mappers/dataset_mapper.py +203 -0
  36. oneformer/data/dataset_mappers/oneformer_unified_dataset_mapper.py +375 -0
  37. oneformer/data/datasets/__init__.py +7 -0
  38. oneformer/data/datasets/register_ade20k_instance.py +56 -0
  39. oneformer/data/datasets/register_ade20k_panoptic.py +394 -0
  40. oneformer/data/datasets/register_cityscapes_panoptic.py +199 -0
  41. oneformer/data/datasets/register_coco_panoptic2instance.py +44 -0
  42. oneformer/data/datasets/register_coco_panoptic_annos_semseg.py +367 -0
  43. oneformer/data/tokenizer.py +193 -0
  44. oneformer/evaluation/__init__.py +3 -0
  45. oneformer/evaluation/cityscapes_evaluation.py +201 -0
  46. oneformer/evaluation/coco_evaluator.py +563 -0
  47. oneformer/evaluation/detection_coco_evaluator.py +723 -0
  48. oneformer/evaluation/evaluator.py +228 -0
  49. oneformer/evaluation/instance_evaluation.py +110 -0
  50. oneformer/modeling/.DS_Store +0 -0
.gitattributes CHANGED
@@ -2,13 +2,13 @@
2
  *.arrow filter=lfs diff=lfs merge=lfs -text
3
  *.bin filter=lfs diff=lfs merge=lfs -text
4
  *.bz2 filter=lfs diff=lfs merge=lfs -text
5
- *.ckpt filter=lfs diff=lfs merge=lfs -text
6
  *.ftz filter=lfs diff=lfs merge=lfs -text
7
  *.gz filter=lfs diff=lfs merge=lfs -text
8
  *.h5 filter=lfs diff=lfs merge=lfs -text
9
  *.joblib filter=lfs diff=lfs merge=lfs -text
 
 
10
  *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
- *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
  *.model filter=lfs diff=lfs merge=lfs -text
13
  *.msgpack filter=lfs diff=lfs merge=lfs -text
14
  *.npy filter=lfs diff=lfs merge=lfs -text
@@ -16,13 +16,13 @@
16
  *.onnx filter=lfs diff=lfs merge=lfs -text
17
  *.ot filter=lfs diff=lfs merge=lfs -text
18
  *.parquet filter=lfs diff=lfs merge=lfs -text
19
- *.pb filter=lfs diff=lfs merge=lfs -text
20
  *.pickle filter=lfs diff=lfs merge=lfs -text
21
  *.pkl filter=lfs diff=lfs merge=lfs -text
 
22
  *.pt filter=lfs diff=lfs merge=lfs -text
 
23
  *.pth filter=lfs diff=lfs merge=lfs -text
24
  *.rar filter=lfs diff=lfs merge=lfs -text
25
- *.safetensors filter=lfs diff=lfs merge=lfs -text
26
  saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
  *.tar.* filter=lfs diff=lfs merge=lfs -text
28
  *.tflite filter=lfs diff=lfs merge=lfs -text
 
2
  *.arrow filter=lfs diff=lfs merge=lfs -text
3
  *.bin filter=lfs diff=lfs merge=lfs -text
4
  *.bz2 filter=lfs diff=lfs merge=lfs -text
 
5
  *.ftz filter=lfs diff=lfs merge=lfs -text
6
  *.gz filter=lfs diff=lfs merge=lfs -text
7
  *.h5 filter=lfs diff=lfs merge=lfs -text
8
  *.joblib filter=lfs diff=lfs merge=lfs -text
9
+ *.jpeg filter=lfs diff=lfs merge=lfs -text
10
+ *.jpg filter=lfs diff=lfs merge=lfs -text
11
  *.lfs.* filter=lfs diff=lfs merge=lfs -text
 
12
  *.model filter=lfs diff=lfs merge=lfs -text
13
  *.msgpack filter=lfs diff=lfs merge=lfs -text
14
  *.npy filter=lfs diff=lfs merge=lfs -text
 
16
  *.onnx filter=lfs diff=lfs merge=lfs -text
17
  *.ot filter=lfs diff=lfs merge=lfs -text
18
  *.parquet filter=lfs diff=lfs merge=lfs -text
 
19
  *.pickle filter=lfs diff=lfs merge=lfs -text
20
  *.pkl filter=lfs diff=lfs merge=lfs -text
21
+ *.pb filter=lfs diff=lfs merge=lfs -text
22
  *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.png filter=lfs diff=lfs merge=lfs -text
24
  *.pth filter=lfs diff=lfs merge=lfs -text
25
  *.rar filter=lfs diff=lfs merge=lfs -text
 
26
  saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
  *.tar.* filter=lfs diff=lfs merge=lfs -text
28
  *.tflite filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,13 +1,14 @@
1
  ---
2
  title: OneFormer
3
- emoji: 🏃
4
- colorFrom: yellow
5
- colorTo: green
6
  sdk: gradio
7
- sdk_version: 3.9.1
8
- app_file: app.py
9
  pinned: false
10
  license: mit
 
11
  ---
12
 
13
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
1
  ---
2
  title: OneFormer
3
+ emoji: 🎗️
4
+ colorFrom: red
5
+ colorTo: blue
6
  sdk: gradio
7
+ sdk_version: 3.1.4
8
+ app_file: gradio_app.py
9
  pinned: false
10
  license: mit
11
+ python_version: 3.8.15
12
  ---
13
 
14
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
configs/.DS_Store ADDED
Binary file (6.15 kB). View file
 
configs/ade20k/Base-ADE20K-UnifiedSegmentation.yaml ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MODEL:
2
+ BACKBONE:
3
+ FREEZE_AT: 0
4
+ NAME: "build_resnet_backbone"
5
+ WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
6
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
7
+ PIXEL_STD: [58.395, 57.120, 57.375]
8
+ RESNETS:
9
+ DEPTH: 50
10
+ STEM_TYPE: "basic" # not used
11
+ STEM_OUT_CHANNELS: 64
12
+ STRIDE_IN_1X1: False
13
+ OUT_FEATURES: ["res2", "res3", "res4", "res5"]
14
+ # NORM: "SyncBN"
15
+ RES5_MULTI_GRID: [1, 1, 1] # not used
16
+ DATASETS:
17
+ TRAIN: ("ade20k_panoptic_train",)
18
+ TEST_PANOPTIC: ("ade20k_panoptic_val",)
19
+ TEST_INSTANCE: ("ade20k_instance_val",)
20
+ TEST_SEMANTIC: ("ade20k_sem_seg_val",)
21
+ SOLVER:
22
+ IMS_PER_BATCH: 16
23
+ BASE_LR: 0.0001
24
+ MAX_ITER: 160000
25
+ WARMUP_FACTOR: 1.0
26
+ WARMUP_ITERS: 0
27
+ WEIGHT_DECAY: 0.05
28
+ OPTIMIZER: "ADAMW"
29
+ LR_SCHEDULER_NAME: "WarmupPolyLR"
30
+ BACKBONE_MULTIPLIER: 0.1
31
+ CLIP_GRADIENTS:
32
+ ENABLED: True
33
+ CLIP_TYPE: "full_model"
34
+ CLIP_VALUE: 0.01
35
+ NORM_TYPE: 2.0
36
+ AMP:
37
+ ENABLED: True
38
+ INPUT:
39
+ MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 512) for x in range(5, 21)]"]
40
+ MIN_SIZE_TRAIN_SAMPLING: "choice"
41
+ MIN_SIZE_TEST: 512
42
+ MAX_SIZE_TRAIN: 2048
43
+ MAX_SIZE_TEST: 2048
44
+ CROP:
45
+ ENABLED: True
46
+ TYPE: "absolute"
47
+ SIZE: (512, 512)
48
+ SINGLE_CATEGORY_MAX_AREA: 1.0
49
+ COLOR_AUG_SSD: True
50
+ SIZE_DIVISIBILITY: 512 # used in dataset mapper
51
+ FORMAT: "RGB"
52
+ DATASET_MAPPER_NAME: "oneformer_unified"
53
+ MAX_SEQ_LEN: 77
54
+ TASK_SEQ_LEN: 77
55
+ TASK_PROB:
56
+ SEMANTIC: 0.33
57
+ INSTANCE: 0.66
58
+ TEST:
59
+ EVAL_PERIOD: 5000
60
+ AUG:
61
+ ENABLED: False
62
+ MIN_SIZES: [256, 384, 512, 640, 768, 896]
63
+ MAX_SIZE: 3584
64
+ FLIP: True
65
+ DATALOADER:
66
+ FILTER_EMPTY_ANNOTATIONS: True
67
+ NUM_WORKERS: 4
68
+ VERSION: 2
configs/ade20k/oneformer_R50_bs16_160k.yaml ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: Base-ADE20K-UnifiedSegmentation.yaml
2
+ MODEL:
3
+ META_ARCHITECTURE: "OneFormer"
4
+ SEM_SEG_HEAD:
5
+ NAME: "OneFormerHead"
6
+ IGNORE_VALUE: 255
7
+ NUM_CLASSES: 150
8
+ LOSS_WEIGHT: 1.0
9
+ CONVS_DIM: 256
10
+ MASK_DIM: 256
11
+ NORM: "GN"
12
+ # pixel decoder
13
+ PIXEL_DECODER_NAME: "MSDeformAttnPixelDecoder"
14
+ IN_FEATURES: ["res2", "res3", "res4", "res5"]
15
+ DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES: ["res3", "res4", "res5"]
16
+ COMMON_STRIDE: 4
17
+ TRANSFORMER_ENC_LAYERS: 6
18
+ ONE_FORMER:
19
+ TRANSFORMER_DECODER_NAME: "ContrastiveMultiScaleMaskedTransformerDecoder"
20
+ TRANSFORMER_IN_FEATURE: "multi_scale_pixel_decoder"
21
+ DEEP_SUPERVISION: True
22
+ NO_OBJECT_WEIGHT: 0.1
23
+ CLASS_WEIGHT: 2.0
24
+ MASK_WEIGHT: 5.0
25
+ DICE_WEIGHT: 5.0
26
+ CONTRASTIVE_WEIGHT: 0.5
27
+ CONTRASTIVE_TEMPERATURE: 0.07
28
+ HIDDEN_DIM: 256
29
+ NUM_OBJECT_QUERIES: 150
30
+ USE_TASK_NORM: True
31
+ NHEADS: 8
32
+ DROPOUT: 0.1
33
+ DIM_FEEDFORWARD: 2048
34
+ ENC_LAYERS: 0
35
+ PRE_NORM: False
36
+ ENFORCE_INPUT_PROJ: False
37
+ SIZE_DIVISIBILITY: 32
38
+ CLASS_DEC_LAYERS: 2
39
+ DEC_LAYERS: 10 # 9 decoder layers, add one for the loss on learnable query
40
+ TRAIN_NUM_POINTS: 12544
41
+ OVERSAMPLE_RATIO: 3.0
42
+ IMPORTANCE_SAMPLE_RATIO: 0.75
43
+ TEXT_ENCODER:
44
+ WIDTH: 256
45
+ CONTEXT_LENGTH: 77
46
+ NUM_LAYERS: 6
47
+ VOCAB_SIZE: 49408
48
+ PROJ_NUM_LAYERS: 2
49
+ N_CTX: 16
50
+ TEST:
51
+ SEMANTIC_ON: True
52
+ INSTANCE_ON: True
53
+ PANOPTIC_ON: True
54
+ OVERLAP_THRESHOLD: 0.8
55
+ OBJECT_MASK_THRESHOLD: 0.8
56
+ TASK: "panoptic"
57
+ TEST:
58
+ DETECTIONS_PER_IMAGE: 150
configs/ade20k/oneformer_dinat_large_IN21k_384_bs16_160k.yaml ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_160k.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2DiNAT"
5
+ DiNAT:
6
+ EMBED_DIM: 192
7
+ MLP_RATIO: 2.0
8
+ DEPTHS: [3, 4, 18, 5]
9
+ NUM_HEADS: [6, 12, 24, 48]
10
+ KERNEL_SIZE: 11
11
+ DROP_PATH_RATE: 0.3
12
+ DILATIONS: [[1, 20, 1], [1, 5, 1, 10], [1, 2, 1, 3, 1, 4, 1, 5, 1, 2, 1, 3, 1, 4, 1, 5, 1, 5], [1, 2, 1, 2, 1]]
13
+ WEIGHTS: "dinat_large_in22k_in1k_384_11x11.pkl"
14
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
15
+ PIXEL_STD: [58.395, 57.120, 57.375]
16
+ ONE_FORMER:
17
+ NUM_OBJECT_QUERIES: 250
18
+ SOLVER:
19
+ AMP:
20
+ ENABLED: False
21
+ INPUT:
22
+ MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 640) for x in range(5, 21)]"]
23
+ MIN_SIZE_TRAIN_SAMPLING: "choice"
24
+ MIN_SIZE_TEST: 640
25
+ MAX_SIZE_TRAIN: 2560
26
+ MAX_SIZE_TEST: 2560
27
+ CROP:
28
+ ENABLED: True
29
+ TYPE: "absolute"
30
+ SIZE: (640, 640)
31
+ SINGLE_CATEGORY_MAX_AREA: 1.0
32
+ COLOR_AUG_SSD: True
33
+ SIZE_DIVISIBILITY: 640 # used in dataset mapper
34
+ FORMAT: "RGB"
35
+ TEST:
36
+ DETECTIONS_PER_IMAGE: 250
37
+ EVAL_PERIOD: 5000
38
+ AUG:
39
+ ENABLED: False
40
+ MIN_SIZES: [320, 480, 640, 800, 960, 1120]
41
+ MAX_SIZE: 4480
42
+ FLIP: True
configs/ade20k/oneformer_swin_large_IN21k_384_bs16_160k.yaml ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_160k.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2SwinTransformer"
5
+ SWIN:
6
+ EMBED_DIM: 192
7
+ DEPTHS: [2, 2, 18, 2]
8
+ NUM_HEADS: [6, 12, 24, 48]
9
+ WINDOW_SIZE: 12
10
+ APE: False
11
+ DROP_PATH_RATE: 0.3
12
+ PATCH_NORM: True
13
+ PRETRAIN_IMG_SIZE: 384
14
+ WEIGHTS: "swin_large_patch4_window12_384_22k.pkl"
15
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
16
+ PIXEL_STD: [58.395, 57.120, 57.375]
17
+ ONE_FORMER:
18
+ NUM_OBJECT_QUERIES: 250
19
+ INPUT:
20
+ MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 640) for x in range(5, 21)]"]
21
+ MIN_SIZE_TRAIN_SAMPLING: "choice"
22
+ MIN_SIZE_TEST: 640
23
+ MAX_SIZE_TRAIN: 2560
24
+ MAX_SIZE_TEST: 2560
25
+ CROP:
26
+ ENABLED: True
27
+ TYPE: "absolute"
28
+ SIZE: (640, 640)
29
+ SINGLE_CATEGORY_MAX_AREA: 1.0
30
+ COLOR_AUG_SSD: True
31
+ SIZE_DIVISIBILITY: 640 # used in dataset mapper
32
+ FORMAT: "RGB"
33
+ TEST:
34
+ DETECTIONS_PER_IMAGE: 250
35
+ EVAL_PERIOD: 5000
36
+ AUG:
37
+ ENABLED: False
38
+ MIN_SIZES: [320, 480, 640, 800, 960, 1120]
39
+ MAX_SIZE: 4480
40
+ FLIP: True
configs/cityscapes/.DS_Store ADDED
Binary file (6.15 kB). View file
 
configs/cityscapes/Base-Cityscapes-UnifiedSegmentation.yaml ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MODEL:
2
+ BACKBONE:
3
+ FREEZE_AT: 0
4
+ NAME: "build_resnet_backbone"
5
+ WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
6
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
7
+ PIXEL_STD: [58.395, 57.120, 57.375]
8
+ RESNETS:
9
+ DEPTH: 50
10
+ STEM_TYPE: "basic" # not used
11
+ STEM_OUT_CHANNELS: 64
12
+ STRIDE_IN_1X1: False
13
+ OUT_FEATURES: ["res2", "res3", "res4", "res5"]
14
+ NORM: "SyncBN" # use syncbn for cityscapes dataset
15
+ RES5_MULTI_GRID: [1, 1, 1] # not used
16
+ DATASETS:
17
+ TRAIN: ("cityscapes_fine_panoptic_train",)
18
+ TEST_PANOPTIC: ("cityscapes_fine_panoptic_val",)
19
+ TEST_INSTANCE: ("cityscapes_fine_instance_seg_val",)
20
+ TEST_SEMANTIC: ("cityscapes_fine_sem_seg_val",)
21
+ SOLVER:
22
+ IMS_PER_BATCH: 16
23
+ BASE_LR: 0.0001
24
+ MAX_ITER: 90000
25
+ WARMUP_FACTOR: 1.0
26
+ WARMUP_ITERS: 0
27
+ WEIGHT_DECAY: 0.05
28
+ OPTIMIZER: "ADAMW"
29
+ LR_SCHEDULER_NAME: "WarmupPolyLR"
30
+ BACKBONE_MULTIPLIER: 0.1
31
+ CLIP_GRADIENTS:
32
+ ENABLED: True
33
+ CLIP_TYPE: "full_model"
34
+ CLIP_VALUE: 0.01
35
+ NORM_TYPE: 2.0
36
+ AMP:
37
+ ENABLED: True
38
+ INPUT:
39
+ MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 1024) for x in range(5, 21)]"]
40
+ MIN_SIZE_TRAIN_SAMPLING: "choice"
41
+ MIN_SIZE_TEST: 1024
42
+ MAX_SIZE_TRAIN: 4096
43
+ MAX_SIZE_TEST: 2048
44
+ CROP:
45
+ ENABLED: True
46
+ TYPE: "absolute"
47
+ SIZE: (512, 1024)
48
+ SINGLE_CATEGORY_MAX_AREA: 1.0
49
+ COLOR_AUG_SSD: True
50
+ SIZE_DIVISIBILITY: -1
51
+ FORMAT: "RGB"
52
+ DATASET_MAPPER_NAME: "oneformer_unified"
53
+ MAX_SEQ_LEN: 77
54
+ TASK_SEQ_LEN: 77
55
+ TASK_PROB:
56
+ SEMANTIC: 0.33
57
+ INSTANCE: 0.66
58
+ TEST:
59
+ EVAL_PERIOD: 5000
60
+ AUG:
61
+ ENABLED: False
62
+ MIN_SIZES: [512, 768, 1024, 1280, 1536, 1792]
63
+ MAX_SIZE: 4096
64
+ FLIP: True
65
+ DATALOADER:
66
+ FILTER_EMPTY_ANNOTATIONS: True
67
+ NUM_WORKERS: 4
68
+ VERSION: 2
configs/cityscapes/oneformer_R50_bs16_90k.yaml ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: Base-Cityscapes-UnifiedSegmentation.yaml
2
+ MODEL:
3
+ META_ARCHITECTURE: "OneFormer"
4
+ SEM_SEG_HEAD:
5
+ NAME: "OneFormerHead"
6
+ IGNORE_VALUE: 255
7
+ NUM_CLASSES: 19
8
+ LOSS_WEIGHT: 1.0
9
+ CONVS_DIM: 256
10
+ MASK_DIM: 256
11
+ NORM: "GN"
12
+ # pixel decoder
13
+ PIXEL_DECODER_NAME: "MSDeformAttnPixelDecoder"
14
+ IN_FEATURES: ["res2", "res3", "res4", "res5"]
15
+ DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES: ["res3", "res4", "res5"]
16
+ COMMON_STRIDE: 4
17
+ TRANSFORMER_ENC_LAYERS: 6
18
+ ONE_FORMER:
19
+ TRANSFORMER_DECODER_NAME: "ContrastiveMultiScaleMaskedTransformerDecoder"
20
+ TRANSFORMER_IN_FEATURE: "multi_scale_pixel_decoder"
21
+ DEEP_SUPERVISION: True
22
+ NO_OBJECT_WEIGHT: 0.1
23
+ CLASS_WEIGHT: 2.0
24
+ MASK_WEIGHT: 5.0
25
+ DICE_WEIGHT: 5.0
26
+ CONTRASTIVE_WEIGHT: 0.5
27
+ CONTRASTIVE_TEMPERATURE: 0.07
28
+ HIDDEN_DIM: 256
29
+ NUM_OBJECT_QUERIES: 150
30
+ USE_TASK_NORM: True
31
+ NHEADS: 8
32
+ DROPOUT: 0.1
33
+ DIM_FEEDFORWARD: 2048
34
+ ENC_LAYERS: 0
35
+ PRE_NORM: False
36
+ ENFORCE_INPUT_PROJ: False
37
+ SIZE_DIVISIBILITY: 32
38
+ ENC_LAYERS: 0
39
+ CLASS_DEC_LAYERS: 2
40
+ DEC_LAYERS: 10 # 9 decoder layers, add one for the loss on learnable query
41
+ TRAIN_NUM_POINTS: 12544
42
+ OVERSAMPLE_RATIO: 3.0
43
+ IMPORTANCE_SAMPLE_RATIO: 0.75
44
+ TEXT_ENCODER:
45
+ WIDTH: 256
46
+ CONTEXT_LENGTH: 77
47
+ NUM_LAYERS: 6
48
+ VOCAB_SIZE: 49408
49
+ PROJ_NUM_LAYERS: 2
50
+ N_CTX: 16
51
+ TEST:
52
+ SEMANTIC_ON: True
53
+ INSTANCE_ON: True
54
+ PANOPTIC_ON: True
55
+ OVERLAP_THRESHOLD: 0.8
56
+ OBJECT_MASK_THRESHOLD: 0.8
57
+ TASK: "panoptic"
58
+ TEST:
59
+ DETECTIONS_PER_IMAGE: 150
configs/cityscapes/oneformer_dinat_large_bs16_90k.yaml ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_90k.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2DiNAT"
5
+ DiNAT:
6
+ EMBED_DIM: 192
7
+ MLP_RATIO: 2.0
8
+ DEPTHS: [3, 4, 18, 5]
9
+ NUM_HEADS: [6, 12, 24, 48]
10
+ KERNEL_SIZE: 7
11
+ DROP_PATH_RATE: 0.3
12
+ DILATIONS: [[1, 18, 1], [1, 5, 1, 9], [1, 2, 1, 3, 1, 4, 1, 2, 1, 3, 1, 4, 1, 2, 1, 3, 1, 4], [1, 2, 1, 2, 1]]
13
+ WEIGHTS: "dinat_large_in22k_224.pkl"
14
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
15
+ PIXEL_STD: [58.395, 57.120, 57.375]
16
+ ONE_FORMER:
17
+ NUM_OBJECT_QUERIES: 250
18
+ SOLVER:
19
+ AMP:
20
+ ENABLED: False
21
+ TEST:
22
+ DETECTIONS_PER_IMAGE: 250
configs/cityscapes/oneformer_swin_large_IN21k_384_bs16_90k.yaml ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_90k.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2SwinTransformer"
5
+ SWIN:
6
+ EMBED_DIM: 192
7
+ DEPTHS: [2, 2, 18, 2]
8
+ NUM_HEADS: [6, 12, 24, 48]
9
+ WINDOW_SIZE: 12
10
+ APE: False
11
+ DROP_PATH_RATE: 0.3
12
+ PATCH_NORM: True
13
+ PRETRAIN_IMG_SIZE: 384
14
+ WEIGHTS: "swin_large_patch4_window12_384_22k.pkl"
15
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
16
+ PIXEL_STD: [58.395, 57.120, 57.375]
17
+ ONE_FORMER:
18
+ NUM_OBJECT_QUERIES: 250
19
+ TEST:
20
+ DETECTIONS_PER_IMAGE: 250
configs/coco/Base-COCO-UnifiedSegmentation.yaml ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MODEL:
2
+ BACKBONE:
3
+ FREEZE_AT: 0
4
+ NAME: "build_resnet_backbone"
5
+ WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
6
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
7
+ PIXEL_STD: [58.395, 57.120, 57.375]
8
+ RESNETS:
9
+ DEPTH: 50
10
+ STEM_TYPE: "basic" # not used
11
+ STEM_OUT_CHANNELS: 64
12
+ STRIDE_IN_1X1: False
13
+ OUT_FEATURES: ["res2", "res3", "res4", "res5"]
14
+ # NORM: "SyncBN"
15
+ RES5_MULTI_GRID: [1, 1, 1] # not used
16
+ DATASETS:
17
+ TRAIN: ("coco_2017_train_panoptic_with_sem_seg",)
18
+ TEST_PANOPTIC: ("coco_2017_val_panoptic_with_sem_seg",) # to evaluate instance and semantic performance as well
19
+ TEST_INSTANCE: ("coco_2017_val",)
20
+ TEST_SEMANTIC: ("coco_2017_val_panoptic_with_sem_seg",)
21
+ SOLVER:
22
+ IMS_PER_BATCH: 16
23
+ BASE_LR: 0.0001
24
+ STEPS: (327778, 355092)
25
+ MAX_ITER: 368750
26
+ WARMUP_FACTOR: 1.0
27
+ WARMUP_ITERS: 10
28
+ WEIGHT_DECAY: 0.05
29
+ OPTIMIZER: "ADAMW"
30
+ BACKBONE_MULTIPLIER: 0.1
31
+ CLIP_GRADIENTS:
32
+ ENABLED: True
33
+ CLIP_TYPE: "full_model"
34
+ CLIP_VALUE: 0.01
35
+ NORM_TYPE: 2.0
36
+ AMP:
37
+ ENABLED: True
38
+ INPUT:
39
+ IMAGE_SIZE: 1024
40
+ MIN_SCALE: 0.1
41
+ MAX_SCALE: 2.0
42
+ FORMAT: "RGB"
43
+ DATASET_MAPPER_NAME: "coco_unified_lsj"
44
+ MAX_SEQ_LEN: 77
45
+ TASK_SEQ_LEN: 77
46
+ TASK_PROB:
47
+ SEMANTIC: 0.33
48
+ INSTANCE: 0.66
49
+ TEST:
50
+ EVAL_PERIOD: 5000
51
+ DATALOADER:
52
+ FILTER_EMPTY_ANNOTATIONS: True
53
+ NUM_WORKERS: 4
54
+ VERSION: 2
configs/coco/oneformer_R50_bs16_50ep.yaml ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: Base-COCO-UnifiedSegmentation.yaml
2
+ MODEL:
3
+ META_ARCHITECTURE: "OneFormer"
4
+ SEM_SEG_HEAD:
5
+ NAME: "OneFormerHead"
6
+ IGNORE_VALUE: 255
7
+ NUM_CLASSES: 133
8
+ LOSS_WEIGHT: 1.0
9
+ CONVS_DIM: 256
10
+ MASK_DIM: 256
11
+ NORM: "GN"
12
+ # pixel decoder
13
+ PIXEL_DECODER_NAME: "MSDeformAttnPixelDecoder"
14
+ IN_FEATURES: ["res2", "res3", "res4", "res5"]
15
+ DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES: ["res3", "res4", "res5"]
16
+ COMMON_STRIDE: 4
17
+ TRANSFORMER_ENC_LAYERS: 6
18
+ ONE_FORMER:
19
+ TRANSFORMER_DECODER_NAME: "ContrastiveMultiScaleMaskedTransformerDecoder"
20
+ TRANSFORMER_IN_FEATURE: "multi_scale_pixel_decoder"
21
+ DEEP_SUPERVISION: True
22
+ NO_OBJECT_WEIGHT: 0.1
23
+ CLASS_WEIGHT: 2.0
24
+ MASK_WEIGHT: 5.0
25
+ DICE_WEIGHT: 5.0
26
+ CONTRASTIVE_WEIGHT: 0.5
27
+ CONTRASTIVE_TEMPERATURE: 0.07
28
+ HIDDEN_DIM: 256
29
+ NUM_OBJECT_QUERIES: 150
30
+ USE_TASK_NORM: True
31
+ NHEADS: 8
32
+ DROPOUT: 0.1
33
+ DIM_FEEDFORWARD: 2048
34
+ ENC_LAYERS: 0
35
+ PRE_NORM: False
36
+ ENFORCE_INPUT_PROJ: False
37
+ SIZE_DIVISIBILITY: 32
38
+ CLASS_DEC_LAYERS: 2
39
+ DEC_LAYERS: 10 # 9 decoder layers, add one for the loss on learnable query
40
+ TRAIN_NUM_POINTS: 12544
41
+ OVERSAMPLE_RATIO: 3.0
42
+ IMPORTANCE_SAMPLE_RATIO: 0.75
43
+ TEXT_ENCODER:
44
+ WIDTH: 256
45
+ CONTEXT_LENGTH: 77
46
+ NUM_LAYERS: 6
47
+ VOCAB_SIZE: 49408
48
+ PROJ_NUM_LAYERS: 2
49
+ N_CTX: 16
50
+ TEST:
51
+ SEMANTIC_ON: True
52
+ INSTANCE_ON: True
53
+ PANOPTIC_ON: True
54
+ DETECTION_ON: False
55
+ OVERLAP_THRESHOLD: 0.8
56
+ OBJECT_MASK_THRESHOLD: 0.8
57
+ TASK: "panoptic"
58
+ TEST:
59
+ DETECTIONS_PER_IMAGE: 150
configs/coco/oneformer_dinat_large_bs16_100ep.yaml ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_50ep.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2DiNAT"
5
+ DiNAT:
6
+ EMBED_DIM: 192
7
+ MLP_RATIO: 2.0
8
+ DEPTHS: [3, 4, 18, 5]
9
+ NUM_HEADS: [6, 12, 24, 48]
10
+ KERNEL_SIZE: 11
11
+ DROP_PATH_RATE: 0.3
12
+ DILATIONS: [[1, 20, 1], [1, 5, 1, 10], [1, 2, 1, 3, 1, 4, 1, 5, 1, 2, 1, 3, 1, 4, 1, 5, 1, 5], [1, 2, 1, 2, 1]]
13
+ WEIGHTS: "dinat_large_in22k_in1k_384_11x11.pkl"
14
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
15
+ PIXEL_STD: [58.395, 57.120, 57.375]
16
+ ONE_FORMER:
17
+ NUM_OBJECT_QUERIES: 150
18
+ SOLVER:
19
+ STEPS: (655556, 710184)
20
+ MAX_ITER: 737500
21
+ TEST:
22
+ DETECTIONS_PER_IMAGE: 150
configs/coco/oneformer_swin_large_IN21k_384_bs16_100ep.yaml ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _BASE_: oneformer_R50_bs16_50ep.yaml
2
+ MODEL:
3
+ BACKBONE:
4
+ NAME: "D2SwinTransformer"
5
+ SWIN:
6
+ EMBED_DIM: 192
7
+ DEPTHS: [2, 2, 18, 2]
8
+ NUM_HEADS: [6, 12, 24, 48]
9
+ WINDOW_SIZE: 12
10
+ APE: False
11
+ DROP_PATH_RATE: 0.3
12
+ PATCH_NORM: True
13
+ PRETRAIN_IMG_SIZE: 384
14
+ WEIGHTS: "swin_large_patch4_window12_384_22k.pkl"
15
+ PIXEL_MEAN: [123.675, 116.280, 103.530]
16
+ PIXEL_STD: [58.395, 57.120, 57.375]
17
+ ONE_FORMER:
18
+ NUM_OBJECT_QUERIES: 150
19
+ SOLVER:
20
+ STEPS: (655556, 735184)
21
+ MAX_ITER: 737500
22
+ AMP:
23
+ ENABLED: False
24
+ TEST:
25
+ DETECTIONS_PER_IMAGE: 150
deform_setup.sh ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ ln -s ./oneformer/modeling/pixel_decoder/ops/ ./
3
+ ls
4
+ cd ops/ && bash make.sh && cd ..
5
+
6
+ echo '----------------------------------------------------------------'
7
+ echo '----------------------------------------------------------------'
8
+ pip3 freeze | grep MultiScaleDeformableAttention
9
+ pip3 freeze | grep torch
10
+ pip3 freeze | grep detectron2
11
+ pip3 freeze | grep natten
12
+ echo '----------------------------------------------------------------'
13
+ echo '----------------------------------------------------------------'
14
+
15
+ echo '----------------------------------------------------------------'
16
+ echo '----------------------------------------------------------------'
17
+ cd /home/user/.pyenv/versions/3.8.15/lib/python3.8/site-packages
18
+ ls
19
+ ls | grep MultiScale
20
+ echo '----------------------------------------------------------------'
21
+ echo '----------------------------------------------------------------'
demo/colormap.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+
3
+ """
4
+ An awesome colormap for really neat visualizations.
5
+ Copied from Detectron, and removed gray colors.
6
+ """
7
+
8
+ import numpy as np
9
+ import random
10
+ random.seed(0)
11
+
12
+ __all__ = ["colormap", "random_color", "random_colors"]
13
+
14
+ # fmt: off
15
+ # RGB:
16
+ # _COLORS = np.array(
17
+ # [
18
+ # 0.000, 0.447, 0.741,
19
+ # 0.850, 0.325, 0.098,
20
+ # 0.929, 0.694, 0.125,
21
+ # 0.494, 0.184, 0.556,
22
+ # 0.466, 0.674, 0.188,
23
+ # 0.301, 0.745, 0.933,
24
+ # 0.635, 0.078, 0.184,
25
+ # 0.300, 0.300, 0.300,
26
+ # 0.600, 0.600, 0.600,
27
+ # 1.000, 0.000, 0.000,
28
+ # 1.000, 0.500, 0.000,
29
+ # 0.749, 0.749, 0.000,
30
+ # 0.000, 1.000, 0.000,
31
+ # 0.000, 0.000, 1.000,
32
+ # 0.667, 0.000, 1.000,
33
+ # 0.333, 0.333, 0.000,
34
+ # 0.333, 0.667, 0.000,
35
+ # 0.333, 1.000, 0.000,
36
+ # 0.667, 0.333, 0.000,
37
+ # 0.667, 0.667, 0.000,
38
+ # 0.667, 1.000, 0.000,
39
+ # 1.000, 0.333, 0.000,
40
+ # 1.000, 0.667, 0.000,
41
+ # 1.000, 1.000, 0.000,
42
+ # 0.000, 0.333, 0.500,
43
+ # 0.000, 0.667, 0.500,
44
+ # 0.000, 1.000, 0.500,
45
+ # 0.333, 0.000, 0.500,
46
+ # 0.333, 0.333, 0.500,
47
+ # 0.333, 0.667, 0.500,
48
+ # 0.333, 1.000, 0.500,
49
+ # 0.667, 0.000, 0.500,
50
+ # 0.667, 0.333, 0.500,
51
+ # 0.667, 0.667, 0.500,
52
+ # 0.667, 1.000, 0.500,
53
+ # 1.000, 0.000, 0.500,
54
+ # 1.000, 0.333, 0.500,
55
+ # 1.000, 0.667, 0.500,
56
+ # 1.000, 1.000, 0.500,
57
+ # 0.000, 0.333, 1.000,
58
+ # 0.000, 0.667, 1.000,
59
+ # 0.000, 1.000, 1.000,
60
+ # 0.333, 0.000, 1.000,
61
+ # 0.333, 0.333, 1.000,
62
+ # 0.333, 0.667, 1.000,
63
+ # 0.333, 1.000, 1.000,
64
+ # 0.667, 0.000, 1.000,
65
+ # 0.667, 0.333, 1.000,
66
+ # 0.667, 0.667, 1.000,
67
+ # 0.667, 1.000, 1.000,
68
+ # 1.000, 0.000, 1.000,
69
+ # 1.000, 0.333, 1.000,
70
+ # 1.000, 0.667, 1.000,
71
+ # 0.333, 0.000, 0.000,
72
+ # 0.500, 0.000, 0.000,
73
+ # 0.667, 0.000, 0.000,
74
+ # 0.833, 0.000, 0.000,
75
+ # 1.000, 0.000, 0.000,
76
+ # 0.000, 0.167, 0.000,
77
+ # 0.000, 0.333, 0.000,
78
+ # 0.000, 0.500, 0.000,
79
+ # 0.000, 0.667, 0.000,
80
+ # 0.000, 0.833, 0.000,
81
+ # 0.000, 1.000, 0.000,
82
+ # 0.000, 0.000, 0.167,
83
+ # 0.000, 0.000, 0.333,
84
+ # 0.000, 0.000, 0.500,
85
+ # 0.000, 0.000, 0.667,
86
+ # 0.000, 0.000, 0.833,
87
+ # 0.000, 0.000, 1.000,
88
+ # 0.000, 0.000, 0.000,
89
+ # 0.143, 0.143, 0.143,
90
+ # 0.857, 0.857, 0.857,
91
+ # 1.000, 1.000, 1.000
92
+ # ]
93
+ # ).astype(np.float32).reshape(-1, 3)
94
+ # fmt: on
95
+
96
+ _COLORS = []
97
+
98
+
99
+ def gen_color():
100
+ color = tuple(np.round(np.random.choice(range(256), size=3)/255, 3))
101
+ if color not in _COLORS and np.mean(color) != 0.0:
102
+ _COLORS.append(color)
103
+ else:
104
+ gen_color()
105
+
106
+
107
+ for _ in range(300):
108
+ gen_color()
109
+
110
+
111
+ def colormap(rgb=False, maximum=255):
112
+ """
113
+ Args:
114
+ rgb (bool): whether to return RGB colors or BGR colors.
115
+ maximum (int): either 255 or 1
116
+ Returns:
117
+ ndarray: a float32 array of Nx3 colors, in range [0, 255] or [0, 1]
118
+ """
119
+ assert maximum in [255, 1], maximum
120
+ c = _COLORS * maximum
121
+ if not rgb:
122
+ c = c[:, ::-1]
123
+ return c
124
+
125
+
126
+ def random_color(rgb=False, maximum=255):
127
+ """
128
+ Args:
129
+ rgb (bool): whether to return RGB colors or BGR colors.
130
+ maximum (int): either 255 or 1
131
+ Returns:
132
+ ndarray: a vector of 3 numbers
133
+ """
134
+ idx = np.random.randint(0, len(_COLORS))
135
+ ret = _COLORS[idx] * maximum
136
+ if not rgb:
137
+ ret = ret[::-1]
138
+ return ret
139
+
140
+
141
+ def random_colors(N, rgb=False, maximum=255):
142
+ """
143
+ Args:
144
+ N (int): number of unique colors needed
145
+ rgb (bool): whether to return RGB colors or BGR colors.
146
+ maximum (int): either 255 or 1
147
+ Returns:
148
+ ndarray: a list of random_color
149
+ """
150
+ indices = random.sample(range(len(_COLORS)), N)
151
+ ret = [_COLORS[i] * maximum for i in indices]
152
+ if not rgb:
153
+ ret = [x[::-1] for x in ret]
154
+ return ret
155
+
156
+
157
+ if __name__ == "__main__":
158
+ import cv2
159
+
160
+ size = 100
161
+ H, W = 10, 10
162
+ canvas = np.random.rand(H * size, W * size, 3).astype("float32")
163
+ for h in range(H):
164
+ for w in range(W):
165
+ idx = h * W + w
166
+ if idx >= len(_COLORS):
167
+ break
168
+ canvas[h * size : (h + 1) * size, w * size : (w + 1) * size] = _COLORS[idx]
169
+ cv2.imshow("a", canvas)
170
+ cv2.waitKey(0)
demo/defaults.py ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import detectron2.data.transforms as T
3
+ from detectron2.checkpoint import DetectionCheckpointer
4
+ from detectron2.data import (
5
+ MetadataCatalog,
6
+ )
7
+ from detectron2.modeling import build_model
8
+
9
+
10
+ __all__ = [
11
+ "DefaultPredictor",
12
+ ]
13
+
14
+
15
+ class DefaultPredictor:
16
+ """
17
+ Create a simple end-to-end predictor with the given config that runs on
18
+ single device for a single input image.
19
+ Compared to using the model directly, this class does the following additions:
20
+ 1. Load checkpoint from `cfg.MODEL.WEIGHTS`.
21
+ 2. Always take BGR image as the input and apply conversion defined by `cfg.INPUT.FORMAT`.
22
+ 3. Apply resizing defined by `cfg.INPUT.{MIN,MAX}_SIZE_TEST`.
23
+ 4. Take one input image and produce a single output, instead of a batch.
24
+ This is meant for simple demo purposes, so it does the above steps automatically.
25
+ This is not meant for benchmarks or running complicated inference logic.
26
+ If you'd like to do anything more complicated, please refer to its source code as
27
+ examples to build and use the model manually.
28
+ Attributes:
29
+ metadata (Metadata): the metadata of the underlying dataset, obtained from
30
+ cfg.DATASETS.TEST.
31
+ Examples:
32
+ ::
33
+ pred = DefaultPredictor(cfg)
34
+ inputs = cv2.imread("input.jpg")
35
+ outputs = pred(inputs)
36
+ """
37
+
38
+ def __init__(self, cfg):
39
+ self.cfg = cfg.clone() # cfg can be modified by model
40
+ self.model = build_model(self.cfg)
41
+ self.model.eval()
42
+ if len(cfg.DATASETS.TEST):
43
+ self.metadata = MetadataCatalog.get(cfg.DATASETS.TEST[0])
44
+
45
+ checkpointer = DetectionCheckpointer(self.model)
46
+ checkpointer.load(cfg.MODEL.WEIGHTS)
47
+
48
+ self.aug = T.ResizeShortestEdge(
49
+ [cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST
50
+ )
51
+
52
+ self.input_format = cfg.INPUT.FORMAT
53
+ assert self.input_format in ["RGB", "BGR"], self.input_format
54
+
55
+ def __call__(self, original_image, task):
56
+ """
57
+ Args:
58
+ original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).
59
+ Returns:
60
+ predictions (dict):
61
+ the output of the model for one image only.
62
+ See :doc:`/tutorials/models` for details about the format.
63
+ """
64
+ with torch.no_grad(): # https://github.com/sphinx-doc/sphinx/issues/4258
65
+ # Apply pre-processing to image.
66
+ if self.input_format == "RGB":
67
+ # whether the model expects BGR inputs or RGB
68
+ original_image = original_image[:, :, ::-1]
69
+ height, width = original_image.shape[:2]
70
+ image = self.aug.get_transform(original_image).apply_image(original_image)
71
+ image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
72
+
73
+ task = f"The task is {task}"
74
+
75
+ inputs = {"image": image, "height": height, "width": width, "task": task}
76
+ predictions = self.model([inputs])[0]
77
+ return predictions
demo/predictor.py ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ # Copied from: https://github.com/facebookresearch/detectron2/blob/master/demo/predictor.py
3
+ import atexit
4
+ import bisect
5
+ import multiprocessing as mp
6
+ from collections import deque
7
+
8
+ import cv2
9
+ import torch
10
+
11
+ from detectron2.data import MetadataCatalog
12
+ from defaults import DefaultPredictor
13
+ from detectron2.utils.video_visualizer import VideoVisualizer
14
+ from visualizer import ColorMode, Visualizer
15
+
16
+
17
+ class VisualizationDemo(object):
18
+ def __init__(self, cfg, instance_mode=ColorMode.IMAGE, parallel=False):
19
+ """
20
+ Args:
21
+ cfg (CfgNode):
22
+ instance_mode (ColorMode):
23
+ parallel (bool): whether to run the model in different processes from visualization.
24
+ Useful since the visualization logic can be slow.
25
+ """
26
+ self.metadata = MetadataCatalog.get(
27
+ cfg.DATASETS.TEST[0] if len(cfg.DATASETS.TEST) else "__unused"
28
+ )
29
+ if 'cityscapes_fine_sem_seg_val' in cfg.DATASETS.TEST[0]:
30
+ from cityscapesscripts.helpers.labels import labels
31
+ stuff_colors = [k.color for k in labels if k.trainId != 255]
32
+ self.metadata = self.metadata.set(stuff_colors=stuff_colors)
33
+ self.cpu_device = torch.device("cpu")
34
+ self.instance_mode = instance_mode
35
+
36
+ self.parallel = parallel
37
+ if parallel:
38
+ num_gpu = torch.cuda.device_count()
39
+ self.predictor = AsyncPredictor(cfg, num_gpus=num_gpu)
40
+ else:
41
+ self.predictor = DefaultPredictor(cfg)
42
+
43
+ def run_on_image(self, image, task, sem_gt, pan_gt, ins_gt, box_gt):
44
+ """
45
+ Args:
46
+ image (np.ndarray): an image of shape (H, W, C) (in BGR order).
47
+ This is the format used by OpenCV.
48
+ Returns:
49
+ predictions (dict): the output of the model.
50
+ vis_output (VisImage): the visualized image output.
51
+ """
52
+ vis_output = None
53
+ # Convert image from OpenCV BGR format to Matplotlib RGB format.
54
+ image = image[:, :, ::-1]
55
+ vis_output = {}
56
+
57
+ if task == 'panoptic':
58
+ visualizer = Visualizer(image, metadata=self.metadata, instance_mode=0)
59
+ predictions = self.predictor(image, "panoptic")
60
+ panoptic_seg, segments_info = predictions["panoptic_seg"]
61
+ vis_output['panoptic'] = visualizer.draw_panoptic_seg_predictions(
62
+ panoptic_seg.to(self.cpu_device), segments_info, alpha=1
63
+ )
64
+
65
+ # visualizer = Visualizer(image, metadata=self.metadata, instance_mode=0)
66
+ # vis_output['pan_gt'] = visualizer.draw_panoptic_seg(
67
+ # pan_gt[0].to(self.cpu_device), pan_gt[1], alpha=1
68
+ # )
69
+
70
+ if task == 'panoptic' or task == 'semantic':
71
+ visualizer = Visualizer(image, metadata=self.metadata, instance_mode=1)
72
+ predictions = self.predictor(image, "semantic")
73
+ vis_output['semantic'] = visualizer.draw_sem_seg(
74
+ predictions["sem_seg"].argmax(dim=0).to(self.cpu_device), alpha=1
75
+ )
76
+
77
+ # visualizer = Visualizer(image, metadata=self.metadata, instance_mode=1)
78
+ # vis_output['gt_sem'] = visualizer.draw_sem_seg(
79
+ # sem_gt.to(self.cpu_device), alpha=1
80
+ # )
81
+
82
+ if task == 'panoptic' or task == 'instance':
83
+ visualizer = Visualizer(image, metadata=self.metadata, instance_mode=2)
84
+ predictions = self.predictor(image, "instance")
85
+ instances = predictions["instances"].to(self.cpu_device)
86
+ vis_output['instance'] = visualizer.draw_instance_predictions(predictions=instances, alpha=1)
87
+
88
+ if 'boxes' in predictions:
89
+ boxes, labels, scores = predictions["boxes"]
90
+ visualizer = Visualizer(image, False, metadata=self.metadata, instance_mode=0)
91
+ vis_output['boxes'] = visualizer.draw_box_predictions(
92
+ boxes.to(self.cpu_device), labels.to(self.cpu_device), scores.to(self.cpu_device))
93
+
94
+
95
+ # visualizer = Visualizer(image, metadata=self.metadata, instance_mode=2)
96
+ # vis_output['ins_gt'] = visualizer.draw_instance_predictions(predictions=ins_gt.to(self.cpu_device), alpha=1)
97
+ # vis_output['input'] = visualizer.get_image(image)
98
+
99
+ return predictions, vis_output
100
+
101
+
102
+ class AsyncPredictor:
103
+ """
104
+ A predictor that runs the model asynchronously, possibly on >1 GPUs.
105
+ Because rendering the visualization takes considerably amount of time,
106
+ this helps improve throughput a little bit when rendering videos.
107
+ """
108
+
109
+ class _StopToken:
110
+ pass
111
+
112
+ class _PredictWorker(mp.Process):
113
+ def __init__(self, cfg, task_queue, result_queue):
114
+ self.cfg = cfg
115
+ self.task_queue = task_queue
116
+ self.result_queue = result_queue
117
+ super().__init__()
118
+
119
+ def run(self):
120
+ predictor = DefaultPredictor(self.cfg)
121
+
122
+ while True:
123
+ task = self.task_queue.get()
124
+ if isinstance(task, AsyncPredictor._StopToken):
125
+ break
126
+ idx, data = task
127
+ result = predictor(data)
128
+ self.result_queue.put((idx, result))
129
+
130
+ def __init__(self, cfg, num_gpus: int = 1):
131
+ """
132
+ Args:
133
+ cfg (CfgNode):
134
+ num_gpus (int): if 0, will run on CPU
135
+ """
136
+ num_workers = max(num_gpus, 1)
137
+ self.task_queue = mp.Queue(maxsize=num_workers * 3)
138
+ self.result_queue = mp.Queue(maxsize=num_workers * 3)
139
+ self.procs = []
140
+ for gpuid in range(max(num_gpus, 1)):
141
+ cfg = cfg.clone()
142
+ cfg.defrost()
143
+ cfg.MODEL.DEVICE = "cuda:{}".format(gpuid) if num_gpus > 0 else "cpu"
144
+ self.procs.append(
145
+ AsyncPredictor._PredictWorker(cfg, self.task_queue, self.result_queue)
146
+ )
147
+
148
+ self.put_idx = 0
149
+ self.get_idx = 0
150
+ self.result_rank = []
151
+ self.result_data = []
152
+
153
+ for p in self.procs:
154
+ p.start()
155
+ atexit.register(self.shutdown)
156
+
157
+ def put(self, image):
158
+ self.put_idx += 1
159
+ self.task_queue.put((self.put_idx, image))
160
+
161
+ def get(self):
162
+ self.get_idx += 1 # the index needed for this request
163
+ if len(self.result_rank) and self.result_rank[0] == self.get_idx:
164
+ res = self.result_data[0]
165
+ del self.result_data[0], self.result_rank[0]
166
+ return res
167
+
168
+ while True:
169
+ # make sure the results are returned in the correct order
170
+ idx, res = self.result_queue.get()
171
+ if idx == self.get_idx:
172
+ return res
173
+ insert = bisect.bisect(self.result_rank, idx)
174
+ self.result_rank.insert(insert, idx)
175
+ self.result_data.insert(insert, res)
176
+
177
+ def __len__(self):
178
+ return self.put_idx - self.get_idx
179
+
180
+ def __call__(self, image):
181
+ self.put(image)
182
+ return self.get()
183
+
184
+ def shutdown(self):
185
+ for _ in self.procs:
186
+ self.task_queue.put(AsyncPredictor._StopToken())
187
+
188
+ @property
189
+ def default_buffer_size(self):
190
+ return len(self.procs) * 5
demo/visualizer.py ADDED
@@ -0,0 +1,1350 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ import colorsys
3
+ import logging
4
+ import math
5
+ import numpy as np
6
+ from enum import Enum, unique
7
+ import cv2
8
+ import matplotlib as mpl
9
+ import matplotlib.colors as mplc
10
+ import matplotlib.figure as mplfigure
11
+ import pycocotools.mask as mask_util
12
+ import torch
13
+ from matplotlib.backends.backend_agg import FigureCanvasAgg
14
+ from PIL import Image
15
+
16
+ from detectron2.data import MetadataCatalog
17
+ from detectron2.structures import BitMasks, Boxes, BoxMode, Keypoints, PolygonMasks, RotatedBoxes
18
+ from detectron2.utils.file_io import PathManager
19
+ import random
20
+ random.seed(0)
21
+ from .colormap import random_color, _COLORS
22
+ logger = logging.getLogger(__name__)
23
+
24
+ __all__ = ["ColorMode", "VisImage", "Visualizer"]
25
+
26
+
27
+ _SMALL_OBJECT_AREA_THRESH = 1000
28
+ _LARGE_MASK_AREA_THRESH = 120000
29
+ _OFF_WHITE = (1.0, 1.0, 240.0 / 255)
30
+ _BLACK = (0, 0, 0)
31
+ _RED = (1.0, 0, 0)
32
+
33
+ _KEYPOINT_THRESHOLD = 0.05
34
+
35
+
36
+ def instance_color(rgb=False, idx=1, maximum=255):
37
+ """
38
+ Args:
39
+ rgb (bool): whether to return RGB colors or BGR colors.
40
+ maximum (int): either 255 or 1
41
+ Returns:
42
+ ndarray: a vector of 3 numbers
43
+ """
44
+ ret = _COLORS[idx] * maximum
45
+ if not rgb:
46
+ ret = ret[::-1]
47
+ return ret
48
+
49
+ @unique
50
+ class ColorMode(Enum):
51
+ """
52
+ Enum of different color modes to use for instance visualizations.
53
+ """
54
+
55
+ IMAGE = 0
56
+ """
57
+ Picks a random color for every instance and overlay segmentations with low opacity.
58
+ """
59
+ SEGMENTATION = 1
60
+ """
61
+ Let instances of the same category have similar colors
62
+ (from metadata.thing_colors), and overlay them with
63
+ high opacity. This provides more attention on the quality of segmentation.
64
+ """
65
+ IMAGE_BW = 2
66
+ """
67
+ Same as IMAGE, but convert all areas without masks to gray-scale.
68
+ Only available for drawing per-instance mask predictions.
69
+ """
70
+
71
+
72
+ class GenericMask:
73
+ """
74
+ Attribute:
75
+ polygons (list[ndarray]): list[ndarray]: polygons for this mask.
76
+ Each ndarray has format [x, y, x, y, ...]
77
+ mask (ndarray): a binary mask
78
+ """
79
+
80
+ def __init__(self, mask_or_polygons, height, width):
81
+ self._mask = self._polygons = self._has_holes = None
82
+ self.height = height
83
+ self.width = width
84
+
85
+ m = mask_or_polygons
86
+ if isinstance(m, dict):
87
+ # RLEs
88
+ assert "counts" in m and "size" in m
89
+ if isinstance(m["counts"], list): # uncompressed RLEs
90
+ h, w = m["size"]
91
+ assert h == height and w == width
92
+ m = mask_util.frPyObjects(m, h, w)
93
+ self._mask = mask_util.decode(m)[:, :]
94
+ return
95
+
96
+ if isinstance(m, list): # list[ndarray]
97
+ self._polygons = [np.asarray(x).reshape(-1) for x in m]
98
+ return
99
+
100
+ if isinstance(m, np.ndarray): # assumed to be a binary mask
101
+ assert m.shape[1] != 2, m.shape
102
+ assert m.shape == (
103
+ height,
104
+ width,
105
+ ), f"mask shape: {m.shape}, target dims: {height}, {width}"
106
+ self._mask = m.astype("uint8")
107
+ return
108
+
109
+ raise ValueError("GenericMask cannot handle object {} of type '{}'".format(m, type(m)))
110
+
111
+ @property
112
+ def mask(self):
113
+ if self._mask is None:
114
+ self._mask = self.polygons_to_mask(self._polygons)
115
+ return self._mask
116
+
117
+ @property
118
+ def polygons(self):
119
+ if self._polygons is None:
120
+ self._polygons, self._has_holes = self.mask_to_polygons(self._mask)
121
+ return self._polygons
122
+
123
+ @property
124
+ def has_holes(self):
125
+ if self._has_holes is None:
126
+ if self._mask is not None:
127
+ self._polygons, self._has_holes = self.mask_to_polygons(self._mask)
128
+ else:
129
+ self._has_holes = False # if original format is polygon, does not have holes
130
+ return self._has_holes
131
+
132
+ def mask_to_polygons(self, mask):
133
+ # cv2.RETR_CCOMP flag retrieves all the contours and arranges them to a 2-level
134
+ # hierarchy. External contours (boundary) of the object are placed in hierarchy-1.
135
+ # Internal contours (holes) are placed in hierarchy-2.
136
+ # cv2.CHAIN_APPROX_NONE flag gets vertices of polygons from contours.
137
+ mask = np.ascontiguousarray(mask) # some versions of cv2 does not support incontiguous arr
138
+ res = cv2.findContours(mask.astype("uint8"), cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE)
139
+ hierarchy = res[-1]
140
+ if hierarchy is None: # empty mask
141
+ return [], False
142
+ has_holes = (hierarchy.reshape(-1, 4)[:, 3] >= 0).sum() > 0
143
+ res = res[-2]
144
+ res = [x.flatten() for x in res]
145
+ # These coordinates from OpenCV are integers in range [0, W-1 or H-1].
146
+ # We add 0.5 to turn them into real-value coordinate space. A better solution
147
+ # would be to first +0.5 and then dilate the returned polygon by 0.5.
148
+ res = [x + 0.5 for x in res if len(x) >= 6]
149
+ return res, has_holes
150
+
151
+ def polygons_to_mask(self, polygons):
152
+ rle = mask_util.frPyObjects(polygons, self.height, self.width)
153
+ rle = mask_util.merge(rle)
154
+ return mask_util.decode(rle)[:, :]
155
+
156
+ def area(self):
157
+ return self.mask.sum()
158
+
159
+ def bbox(self):
160
+ p = mask_util.frPyObjects(self.polygons, self.height, self.width)
161
+ p = mask_util.merge(p)
162
+ bbox = mask_util.toBbox(p)
163
+ bbox[2] += bbox[0]
164
+ bbox[3] += bbox[1]
165
+ return bbox
166
+
167
+
168
+ class _PanopticPrediction:
169
+ """
170
+ Unify different panoptic annotation/prediction formats
171
+ """
172
+
173
+ def __init__(self, panoptic_seg, segments_info, metadata=None):
174
+ if segments_info is None:
175
+ assert metadata is not None
176
+ # If "segments_info" is None, we assume "panoptic_img" is a
177
+ # H*W int32 image storing the panoptic_id in the format of
178
+ # category_id * label_divisor + instance_id. We reserve -1 for
179
+ # VOID label.
180
+ label_divisor = metadata.label_divisor
181
+ segments_info = []
182
+ for panoptic_label in np.unique(panoptic_seg.numpy()):
183
+ if panoptic_label == -1:
184
+ # VOID region.
185
+ continue
186
+ pred_class = panoptic_label // label_divisor
187
+ isthing = pred_class in metadata.thing_dataset_id_to_contiguous_id.values()
188
+ segments_info.append(
189
+ {
190
+ "id": int(panoptic_label),
191
+ "category_id": int(pred_class),
192
+ "isthing": bool(isthing),
193
+ }
194
+ )
195
+ del metadata
196
+
197
+ self._seg = panoptic_seg
198
+
199
+ self._sinfo = {s["id"]: s for s in segments_info} # seg id -> seg info
200
+ segment_ids, areas = torch.unique(panoptic_seg, sorted=True, return_counts=True)
201
+ areas = areas.numpy()
202
+ sorted_idxs = np.argsort(-areas)
203
+ self._seg_ids, self._seg_areas = segment_ids[sorted_idxs], areas[sorted_idxs]
204
+ self._seg_ids = self._seg_ids.tolist()
205
+ for sid, area in zip(self._seg_ids, self._seg_areas):
206
+ if sid in self._sinfo:
207
+ self._sinfo[sid]["area"] = float(area)
208
+
209
+ def non_empty_mask(self):
210
+ """
211
+ Returns:
212
+ (H, W) array, a mask for all pixels that have a prediction
213
+ """
214
+ empty_ids = []
215
+ for id in self._seg_ids:
216
+ if id not in self._sinfo:
217
+ empty_ids.append(id)
218
+ if len(empty_ids) == 0:
219
+ return np.zeros(self._seg.shape, dtype=np.uint8)
220
+ assert (
221
+ len(empty_ids) == 1
222
+ ), ">1 ids corresponds to no labels. This is currently not supported"
223
+ return (self._seg != empty_ids[0]).numpy().astype(np.bool)
224
+
225
+ def semantic_masks(self):
226
+ for sid in self._seg_ids:
227
+ sinfo = self._sinfo.get(sid)
228
+ if sinfo is None or sinfo["isthing"]:
229
+ # Some pixels (e.g. id 0 in PanopticFPN) have no instance or semantic predictions.
230
+ continue
231
+ yield (self._seg == sid).numpy().astype(np.bool), sinfo
232
+
233
+ def instance_masks(self):
234
+ for sid in self._seg_ids:
235
+ sinfo = self._sinfo.get(sid)
236
+ if sinfo is None or not sinfo["isthing"]:
237
+ continue
238
+ mask = (self._seg == sid).numpy().astype(np.bool)
239
+ if mask.sum() > 0:
240
+ yield mask, sinfo
241
+
242
+
243
+ def _create_text_labels(classes, scores, class_names, is_crowd=None):
244
+ """
245
+ Args:
246
+ classes (list[int] or None):
247
+ scores (list[float] or None):
248
+ class_names (list[str] or None):
249
+ is_crowd (list[bool] or None):
250
+ Returns:
251
+ list[str] or None
252
+ """
253
+ labels = None
254
+ if classes is not None:
255
+ if class_names is not None and len(class_names) > 0:
256
+ labels = [class_names[i] for i in classes]
257
+ else:
258
+ labels = [str(i) for i in classes]
259
+ if scores is not None:
260
+ if labels is None:
261
+ labels = ["{:.0f}%".format(s * 100) for s in scores]
262
+ else:
263
+ labels = ["{} {:.0f}%".format(l, s * 100) for l, s in zip(labels, scores)]
264
+ if labels is not None and is_crowd is not None:
265
+ labels = [l + ("|crowd" if crowd else "") for l, crowd in zip(labels, is_crowd)]
266
+ return labels
267
+
268
+
269
+ class VisImage:
270
+ def __init__(self, img, scale=1.0):
271
+ """
272
+ Args:
273
+ img (ndarray): an RGB image of shape (H, W, 3) in range [0, 255].
274
+ scale (float): scale the input image
275
+ """
276
+ self.img = img
277
+ self.scale = scale
278
+ self.width, self.height = img.shape[1], img.shape[0]
279
+ self._setup_figure(img)
280
+
281
+ def _setup_figure(self, img):
282
+ """
283
+ Args:
284
+ Same as in :meth:`__init__()`.
285
+ Returns:
286
+ fig (matplotlib.pyplot.figure): top level container for all the image plot elements.
287
+ ax (matplotlib.pyplot.Axes): contains figure elements and sets the coordinate system.
288
+ """
289
+ fig = mplfigure.Figure(frameon=False)
290
+ self.dpi = fig.get_dpi()
291
+ # add a small 1e-2 to avoid precision lost due to matplotlib's truncation
292
+ # (https://github.com/matplotlib/matplotlib/issues/15363)
293
+ fig.set_size_inches(
294
+ (self.width * self.scale + 1e-2) / self.dpi,
295
+ (self.height * self.scale + 1e-2) / self.dpi,
296
+ )
297
+ self.canvas = FigureCanvasAgg(fig)
298
+ # self.canvas = mpl.backends.backend_cairo.FigureCanvasCairo(fig)
299
+ ax = fig.add_axes([0.0, 0.0, 1.0, 1.0])
300
+ ax.axis("off")
301
+ self.fig = fig
302
+ self.ax = ax
303
+ self.reset_image(img)
304
+
305
+ def reset_image(self, img):
306
+ """
307
+ Args:
308
+ img: same as in __init__
309
+ """
310
+ img = img.astype("uint8")
311
+ self.ax.imshow(img, extent=(0, self.width, self.height, 0), interpolation="nearest")
312
+
313
+ def save(self, filepath):
314
+ """
315
+ Args:
316
+ filepath (str): a string that contains the absolute path, including the file name, where
317
+ the visualized image will be saved.
318
+ """
319
+ self.fig.savefig(filepath)
320
+
321
+ def get_image(self):
322
+ """
323
+ Returns:
324
+ ndarray:
325
+ the visualized image of shape (H, W, 3) (RGB) in uint8 type.
326
+ The shape is scaled w.r.t the input image using the given `scale` argument.
327
+ """
328
+ canvas = self.canvas
329
+ s, (width, height) = canvas.print_to_buffer()
330
+ # buf = io.BytesIO() # works for cairo backend
331
+ # canvas.print_rgba(buf)
332
+ # width, height = self.width, self.height
333
+ # s = buf.getvalue()
334
+
335
+ buffer = np.frombuffer(s, dtype="uint8")
336
+
337
+ img_rgba = buffer.reshape(height, width, 4)
338
+ rgb, alpha = np.split(img_rgba, [3], axis=2)
339
+ return rgb.astype("uint8")
340
+
341
+
342
+ class Visualizer:
343
+ """
344
+ Visualizer that draws data about detection/segmentation on images.
345
+ It contains methods like `draw_{text,box,circle,line,binary_mask,polygon}`
346
+ that draw primitive objects to images, as well as high-level wrappers like
347
+ `draw_{instance_predictions,sem_seg,panoptic_seg_predictions,dataset_dict}`
348
+ that draw composite data in some pre-defined style.
349
+ Note that the exact visualization style for the high-level wrappers are subject to change.
350
+ Style such as color, opacity, label contents, visibility of labels, or even the visibility
351
+ of objects themselves (e.g. when the object is too small) may change according
352
+ to different heuristics, as long as the results still look visually reasonable.
353
+ To obtain a consistent style, you can implement custom drawing functions with the
354
+ abovementioned primitive methods instead. If you need more customized visualization
355
+ styles, you can process the data yourself following their format documented in
356
+ tutorials (:doc:`/tutorials/models`, :doc:`/tutorials/datasets`). This class does not
357
+ intend to satisfy everyone's preference on drawing styles.
358
+ This visualizer focuses on high rendering quality rather than performance. It is not
359
+ designed to be used for real-time applications.
360
+ """
361
+
362
+ # TODO implement a fast, rasterized version using OpenCV
363
+
364
+ def __init__(self, img_rgb, is_img=True, metadata=None, scale=1.0, instance_mode=ColorMode.IMAGE):
365
+ """
366
+ Args:
367
+ img_rgb: a numpy array of shape (H, W, C), where H and W correspond to
368
+ the height and width of the image respectively. C is the number of
369
+ color channels. The image is required to be in RGB format since that
370
+ is a requirement of the Matplotlib library. The image is also expected
371
+ to be in the range [0, 255].
372
+ metadata (Metadata): dataset metadata (e.g. class names and colors)
373
+ instance_mode (ColorMode): defines one of the pre-defined style for drawing
374
+ instances on an image.
375
+ """
376
+ if is_img:
377
+ self.img = np.asarray(img_rgb).clip(0, 255).astype(np.uint8)
378
+ else:
379
+ self.img = np.zeros_like(img_rgb).clip(0, 255).astype(np.uint8)
380
+ if metadata is None:
381
+ metadata = MetadataCatalog.get("__nonexist__")
382
+ self.metadata = metadata
383
+ self.output = VisImage(self.img, scale=scale)
384
+ self.cpu_device = torch.device("cpu")
385
+
386
+ # too small texts are useless, therefore clamp to 9
387
+ self._default_font_size = max(
388
+ np.sqrt(self.output.height * self.output.width) // 90, 10 // scale
389
+ )
390
+ self._instance_mode = instance_mode
391
+ self.keypoint_threshold = _KEYPOINT_THRESHOLD
392
+
393
+ def get_image(self, img):
394
+ img = np.asarray(img).clip(0, 255).astype(np.uint8)
395
+ return VisImage(img, scale=1.0)
396
+
397
+ def draw_box_predictions(
398
+ self,
399
+ boxes=None,
400
+ labels=None,
401
+ scores=None,
402
+ assigned_colors=None
403
+ ):
404
+ """
405
+ Args:
406
+ boxes (Boxes, RotatedBoxes or ndarray): either a :class:`Boxes`,
407
+ or an Nx4 numpy array of XYXY_ABS format for the N objects in a single image,
408
+ or a :class:`RotatedBoxes`,
409
+ or an Nx5 numpy array of (x_center, y_center, width, height, angle_degrees) format
410
+ for the N objects in a single image,
411
+ labels (list[str]): the text to be displayed for each instance.
412
+ assigned_colors (list[matplotlib.colors]): a list of colors, where each color
413
+ corresponds to each mask or box in the image. Refer to 'matplotlib.colors'
414
+ for full list of formats that the colors are accepted in.
415
+ Returns:
416
+ output (VisImage): image object with visualizations.
417
+ """
418
+ num_instances = 0
419
+ boxes = self._convert_boxes(boxes)
420
+ classes = labels.tolist()
421
+ scores = scores.tolist()
422
+ labels = _create_text_labels(classes, scores, self.metadata.get("stuff_classes", None))
423
+ num_instances = len(boxes)
424
+ assert len(labels) == num_instances
425
+ if assigned_colors is None:
426
+ # assigned_colors = [random_color(rgb=True, maximum=1) for _ in range(num_instances)]
427
+ assigned_colors = [instance_color(rgb=True, idx=i, maximum=1) for i in range(num_instances)]
428
+ if num_instances == 0:
429
+ return self.output
430
+
431
+ # Display in largest to smallest order to reduce occlusion.
432
+ areas = None
433
+ areas = np.prod(boxes[:, 2:] - boxes[:, :2], axis=1)
434
+
435
+ if areas is not None:
436
+ sorted_idxs = np.argsort(-areas).tolist()
437
+ # Re-order overlapped instances in descending order.
438
+ boxes = boxes[sorted_idxs] if boxes is not None else None
439
+ labels = [labels[k] for k in sorted_idxs] if labels is not None else None
440
+ assigned_colors = [assigned_colors[idx] for idx in sorted_idxs]
441
+
442
+ for i in range(num_instances):
443
+ color = assigned_colors[i]
444
+ if boxes is not None:
445
+ self.draw_box(boxes[i], edge_color=color)
446
+
447
+ if labels is not None:
448
+ # first get a box
449
+ if boxes is not None:
450
+ x0, y0, x1, y1 = boxes[i]
451
+ text_pos = (x0, y0) # if drawing boxes, put text on the box corner.
452
+ horiz_align = "left"
453
+ else:
454
+ continue # drawing the box confidence for keypoints isn't very useful.
455
+ # for small objects, draw text at the side to avoid occlusion
456
+ instance_area = (y1 - y0) * (x1 - x0)
457
+ if (
458
+ instance_area < _SMALL_OBJECT_AREA_THRESH * self.output.scale
459
+ or y1 - y0 < 40 * self.output.scale
460
+ ):
461
+ if y1 >= self.output.height - 5:
462
+ text_pos = (x1, y0)
463
+ else:
464
+ text_pos = (x0, y1)
465
+
466
+ height_ratio = (y1 - y0) / np.sqrt(self.output.height * self.output.width)
467
+ lighter_color = self._change_color_brightness(color, brightness_factor=0.7)
468
+ font_size = (
469
+ np.clip((height_ratio - 0.02) / 0.08 + 1, 1.2, 2)
470
+ * 0.5
471
+ * self._default_font_size
472
+ )
473
+ self.draw_text(
474
+ labels[i],
475
+ text_pos,
476
+ color=lighter_color,
477
+ horizontal_alignment=horiz_align,
478
+ font_size=font_size,
479
+ )
480
+
481
+ return self.output
482
+
483
+
484
+ def draw_instance_predictions(self, predictions, alpha=0.8, is_text=True):
485
+ """
486
+ Draw instance-level prediction results on an image.
487
+ Args:
488
+ predictions (Instances): the output of an instance detection/segmentation
489
+ model. Following fields will be used to draw:
490
+ "pred_boxes", "pred_classes", "scores", "pred_masks" (or "pred_masks_rle").
491
+ Returns:
492
+ output (VisImage): image object with visualizations.
493
+ """
494
+ boxes = predictions.pred_boxes if predictions.has("pred_boxes") else None
495
+ scores = predictions.scores if predictions.has("scores") else None
496
+ classes = predictions.pred_classes.tolist() if predictions.has("pred_classes") else None
497
+ labels = _create_text_labels(classes, scores, self.metadata.get("stuff_classes", None))
498
+ keypoints = predictions.pred_keypoints if predictions.has("pred_keypoints") else None
499
+
500
+ if predictions.has("pred_masks"):
501
+ masks = np.asarray(predictions.pred_masks)
502
+ masks = [GenericMask(x, self.output.height, self.output.width) for x in masks]
503
+ else:
504
+ masks = None
505
+
506
+ if self._instance_mode == ColorMode.SEGMENTATION and self.metadata.get("stuff_colors"):
507
+ # colors = [
508
+ # self._jitter([x / 255 for x in self.metadata.thing_colors[c]]) for c in classes
509
+ # ]
510
+ colors = [
511
+ instance_color(rgb=True, idx=c, maximum=1) for c in classes
512
+ ]
513
+ else:
514
+ colors = None
515
+
516
+ if self._instance_mode == ColorMode.IMAGE_BW:
517
+ self.output.reset_image(
518
+ self._create_grayscale_image(
519
+ (predictions.pred_masks.any(dim=0) > 0).numpy()
520
+ if predictions.has("pred_masks")
521
+ else None
522
+ )
523
+ )
524
+
525
+ self.overlay_instances(
526
+ masks=masks,
527
+ boxes=boxes,
528
+ labels=labels,
529
+ keypoints=keypoints,
530
+ assigned_colors=colors,
531
+ alpha=alpha,
532
+ is_text=is_text,
533
+ )
534
+ return self.output
535
+
536
+ def draw_sem_seg(self, sem_seg, area_threshold=None, alpha=0.8, is_text=True):
537
+ """
538
+ Draw semantic segmentation predictions/labels.
539
+ Args:
540
+ sem_seg (Tensor or ndarray): the segmentation of shape (H, W).
541
+ Each value is the integer label of the pixel.
542
+ area_threshold (int): segments with less than `area_threshold` are not drawn.
543
+ alpha (float): the larger it is, the more opaque the segmentations are.
544
+ Returns:
545
+ output (VisImage): image object with visualizations.
546
+ """
547
+ if isinstance(sem_seg, torch.Tensor):
548
+ sem_seg = sem_seg.numpy()
549
+ labels, areas = np.unique(sem_seg, return_counts=True)
550
+ sorted_idxs = np.argsort(-areas).tolist()
551
+ labels = labels[sorted_idxs]
552
+ for label in filter(lambda l: l < len(self.metadata.stuff_classes), labels):
553
+ try:
554
+ mask_color = [x / 255 for x in self.metadata.stuff_colors[label]]
555
+ except (AttributeError, IndexError):
556
+ mask_color = None
557
+
558
+ binary_mask = (sem_seg == label).astype(np.uint8)
559
+ text = self.metadata.stuff_classes[label]
560
+ self.draw_binary_mask(
561
+ binary_mask,
562
+ color=mask_color,
563
+ edge_color=_OFF_WHITE,
564
+ text=text,
565
+ alpha=alpha,
566
+ area_threshold=area_threshold,
567
+ is_text=is_text,
568
+ )
569
+ return self.output
570
+
571
+ def draw_panoptic_seg(self, panoptic_seg, segments_info, area_threshold=None, alpha=0.7, is_text=True,):
572
+ """
573
+ Draw panoptic prediction annotations or results.
574
+ Args:
575
+ panoptic_seg (Tensor): of shape (height, width) where the values are ids for each
576
+ segment.
577
+ segments_info (list[dict] or None): Describe each segment in `panoptic_seg`.
578
+ If it is a ``list[dict]``, each dict contains keys "id", "category_id".
579
+ If None, category id of each pixel is computed by
580
+ ``pixel // metadata.label_divisor``.
581
+ area_threshold (int): stuff segments with less than `area_threshold` are not drawn.
582
+ Returns:
583
+ output (VisImage): image object with visualizations.
584
+ """
585
+ pred = _PanopticPrediction(panoptic_seg, segments_info, self.metadata)
586
+
587
+ if self._instance_mode == ColorMode.IMAGE_BW:
588
+ self.output.reset_image(self._create_grayscale_image(pred.non_empty_mask()))
589
+
590
+ # draw mask for all semantic segments first i.e. "stuff"
591
+ for mask, sinfo in pred.semantic_masks():
592
+ category_idx = sinfo["category_id"]
593
+ try:
594
+ mask_color = [x / 255 for x in self.metadata.stuff_colors[category_idx]]
595
+ except AttributeError:
596
+ mask_color = None
597
+
598
+ text = self.metadata.stuff_classes[category_idx]
599
+ self.draw_binary_mask(
600
+ mask,
601
+ color=mask_color,
602
+ edge_color=_OFF_WHITE,
603
+ text=text,
604
+ alpha=alpha,
605
+ area_threshold=area_threshold,
606
+ is_text=is_text,
607
+ )
608
+
609
+ # draw mask for all instances second
610
+ all_instances = list(pred.instance_masks())
611
+ if len(all_instances) == 0:
612
+ return self.output
613
+ masks, sinfo = list(zip(*all_instances))
614
+ category_ids = [x["category_id"] for x in sinfo]
615
+
616
+ try:
617
+ scores = [x["score"] for x in sinfo]
618
+ except KeyError:
619
+ scores = None
620
+ labels = _create_text_labels(
621
+ category_ids, scores, self.metadata.stuff_classes, [x.get("iscrowd", 0) for x in sinfo]
622
+ )
623
+
624
+ try:
625
+ colors = [
626
+ self._jitter([x / 255 for x in self.metadata.stuff_colors[c]]) for c in category_ids
627
+ ]
628
+ except AttributeError:
629
+ colors = None
630
+ self.overlay_instances(masks=masks, labels=labels, assigned_colors=colors, alpha=alpha, is_text=is_text)
631
+
632
+ return self.output
633
+
634
+ draw_panoptic_seg_predictions = draw_panoptic_seg # backward compatibility
635
+
636
+ def draw_dataset_dict(self, dic):
637
+ """
638
+ Draw annotations/segmentaions in Detectron2 Dataset format.
639
+ Args:
640
+ dic (dict): annotation/segmentation data of one image, in Detectron2 Dataset format.
641
+ Returns:
642
+ output (VisImage): image object with visualizations.
643
+ """
644
+ annos = dic.get("annotations", None)
645
+ if annos:
646
+ if "segmentation" in annos[0]:
647
+ masks = [x["segmentation"] for x in annos]
648
+ else:
649
+ masks = None
650
+ if "keypoints" in annos[0]:
651
+ keypts = [x["keypoints"] for x in annos]
652
+ keypts = np.array(keypts).reshape(len(annos), -1, 3)
653
+ else:
654
+ keypts = None
655
+
656
+ boxes = [
657
+ BoxMode.convert(x["bbox"], x["bbox_mode"], BoxMode.XYXY_ABS)
658
+ if len(x["bbox"]) == 4
659
+ else x["bbox"]
660
+ for x in annos
661
+ ]
662
+
663
+ colors = None
664
+ category_ids = [x["category_id"] for x in annos]
665
+ if self._instance_mode == ColorMode.SEGMENTATION and self.metadata.get("stuff_colors"):
666
+ colors = [
667
+ self._jitter([x / 255 for x in self.metadata.stuff_colors[c]])
668
+ for c in category_ids
669
+ ]
670
+ names = self.metadata.get("stuff_classes", None)
671
+ labels = _create_text_labels(
672
+ category_ids,
673
+ scores=None,
674
+ class_names=names,
675
+ is_crowd=[x.get("iscrowd", 0) for x in annos],
676
+ )
677
+ self.overlay_instances(
678
+ labels=labels, boxes=boxes, masks=masks, keypoints=keypts, assigned_colors=colors
679
+ )
680
+
681
+ sem_seg = dic.get("sem_seg", None)
682
+ if sem_seg is None and "sem_seg_file_name" in dic:
683
+ with PathManager.open(dic["sem_seg_file_name"], "rb") as f:
684
+ sem_seg = Image.open(f)
685
+ sem_seg = np.asarray(sem_seg, dtype="uint8")
686
+ if sem_seg is not None:
687
+ self.draw_sem_seg(sem_seg, area_threshold=0, alpha=0.5)
688
+
689
+ pan_seg = dic.get("pan_seg", None)
690
+ if pan_seg is None and "pan_seg_file_name" in dic:
691
+ with PathManager.open(dic["pan_seg_file_name"], "rb") as f:
692
+ pan_seg = Image.open(f)
693
+ pan_seg = np.asarray(pan_seg)
694
+ from panopticapi.utils import rgb2id
695
+
696
+ pan_seg = rgb2id(pan_seg)
697
+ if pan_seg is not None:
698
+ segments_info = dic["segments_info"]
699
+ pan_seg = torch.tensor(pan_seg)
700
+ self.draw_panoptic_seg(pan_seg, segments_info, area_threshold=0, alpha=0.5)
701
+ return self.output
702
+
703
+ def overlay_instances(
704
+ self,
705
+ *,
706
+ boxes=None,
707
+ labels=None,
708
+ masks=None,
709
+ keypoints=None,
710
+ assigned_colors=None,
711
+ alpha=0.5,
712
+ is_text=True,
713
+ ):
714
+ """
715
+ Args:
716
+ boxes (Boxes, RotatedBoxes or ndarray): either a :class:`Boxes`,
717
+ or an Nx4 numpy array of XYXY_ABS format for the N objects in a single image,
718
+ or a :class:`RotatedBoxes`,
719
+ or an Nx5 numpy array of (x_center, y_center, width, height, angle_degrees) format
720
+ for the N objects in a single image,
721
+ labels (list[str]): the text to be displayed for each instance.
722
+ masks (masks-like object): Supported types are:
723
+ * :class:`detectron2.structures.PolygonMasks`,
724
+ :class:`detectron2.structures.BitMasks`.
725
+ * list[list[ndarray]]: contains the segmentation masks for all objects in one image.
726
+ The first level of the list corresponds to individual instances. The second
727
+ level to all the polygon that compose the instance, and the third level
728
+ to the polygon coordinates. The third level should have the format of
729
+ [x0, y0, x1, y1, ..., xn, yn] (n >= 3).
730
+ * list[ndarray]: each ndarray is a binary mask of shape (H, W).
731
+ * list[dict]: each dict is a COCO-style RLE.
732
+ keypoints (Keypoint or array like): an array-like object of shape (N, K, 3),
733
+ where the N is the number of instances and K is the number of keypoints.
734
+ The last dimension corresponds to (x, y, visibility or score).
735
+ assigned_colors (list[matplotlib.colors]): a list of colors, where each color
736
+ corresponds to each mask or box in the image. Refer to 'matplotlib.colors'
737
+ for full list of formats that the colors are accepted in.
738
+ Returns:
739
+ output (VisImage): image object with visualizations.
740
+ """
741
+ num_instances = 0
742
+ if boxes is not None:
743
+ boxes = self._convert_boxes(boxes)
744
+ num_instances = len(boxes)
745
+ if masks is not None:
746
+ masks = self._convert_masks(masks)
747
+ if num_instances:
748
+ assert len(masks) == num_instances
749
+ else:
750
+ num_instances = len(masks)
751
+ if keypoints is not None:
752
+ if num_instances:
753
+ assert len(keypoints) == num_instances
754
+ else:
755
+ num_instances = len(keypoints)
756
+ keypoints = self._convert_keypoints(keypoints)
757
+ if labels is not None:
758
+ assert len(labels) == num_instances
759
+ if assigned_colors is None:
760
+ # assigned_colors = [random_color(rgb=True, maximum=1) for _ in range(num_instances)]
761
+ assigned_colors = [instance_color(rgb=True, idx=i, maximum=1) for i in range(num_instances)]
762
+ if num_instances == 0:
763
+ return self.output
764
+ if boxes is not None and boxes.shape[1] == 5:
765
+ return self.overlay_rotated_instances(
766
+ boxes=boxes, labels=labels, assigned_colors=assigned_colors
767
+ )
768
+
769
+ # Display in largest to smallest order to reduce occlusion.
770
+ areas = None
771
+ if boxes is not None:
772
+ areas = np.prod(boxes[:, 2:] - boxes[:, :2], axis=1)
773
+ elif masks is not None:
774
+ areas = np.asarray([x.area() for x in masks])
775
+
776
+ if areas is not None:
777
+ sorted_idxs = np.argsort(-areas).tolist()
778
+ # Re-order overlapped instances in descending order.
779
+ boxes = boxes[sorted_idxs] if boxes is not None else None
780
+ labels = [labels[k] for k in sorted_idxs] if labels is not None else None
781
+ masks = [masks[idx] for idx in sorted_idxs] if masks is not None else None
782
+ assigned_colors = [assigned_colors[idx] for idx in sorted_idxs]
783
+ keypoints = keypoints[sorted_idxs] if keypoints is not None else None
784
+
785
+ for i in range(num_instances):
786
+ color = assigned_colors[i]
787
+ if boxes is not None:
788
+ self.draw_box(boxes[i], edge_color=color)
789
+
790
+ if masks is not None:
791
+ for segment in masks[i].polygons:
792
+ self.draw_polygon(segment.reshape(-1, 2), color, alpha=alpha)
793
+
794
+ if labels is not None:
795
+ # first get a box
796
+ if boxes is not None:
797
+ x0, y0, x1, y1 = boxes[i]
798
+ text_pos = (x0, y0) # if drawing boxes, put text on the box corner.
799
+ horiz_align = "left"
800
+ elif masks is not None:
801
+ # skip small mask without polygon
802
+ if len(masks[i].polygons) == 0:
803
+ continue
804
+
805
+ x0, y0, x1, y1 = masks[i].bbox()
806
+
807
+ # draw text in the center (defined by median) when box is not drawn
808
+ # median is less sensitive to outliers.
809
+ text_pos = np.median(masks[i].mask.nonzero(), axis=1)[::-1]
810
+ horiz_align = "center"
811
+ else:
812
+ continue # drawing the box confidence for keypoints isn't very useful.
813
+ # for small objects, draw text at the side to avoid occlusion
814
+ instance_area = (y1 - y0) * (x1 - x0)
815
+ if (
816
+ instance_area < _SMALL_OBJECT_AREA_THRESH * self.output.scale
817
+ or y1 - y0 < 40 * self.output.scale
818
+ ):
819
+ if y1 >= self.output.height - 5:
820
+ text_pos = (x1, y0)
821
+ else:
822
+ text_pos = (x0, y1)
823
+
824
+ height_ratio = (y1 - y0) / np.sqrt(self.output.height * self.output.width)
825
+ lighter_color = self._change_color_brightness(color, brightness_factor=0.7)
826
+ font_size = (
827
+ np.clip((height_ratio - 0.02) / 0.08 + 1, 1.2, 2)
828
+ * 0.5
829
+ * self._default_font_size
830
+ )
831
+ if is_text:
832
+ self.draw_text(
833
+ labels[i],
834
+ text_pos,
835
+ color=lighter_color,
836
+ horizontal_alignment=horiz_align,
837
+ font_size=font_size,
838
+ )
839
+
840
+ # draw keypoints
841
+ if keypoints is not None:
842
+ for keypoints_per_instance in keypoints:
843
+ self.draw_and_connect_keypoints(keypoints_per_instance)
844
+
845
+ return self.output
846
+
847
+ def overlay_rotated_instances(self, boxes=None, labels=None, assigned_colors=None):
848
+ """
849
+ Args:
850
+ boxes (ndarray): an Nx5 numpy array of
851
+ (x_center, y_center, width, height, angle_degrees) format
852
+ for the N objects in a single image.
853
+ labels (list[str]): the text to be displayed for each instance.
854
+ assigned_colors (list[matplotlib.colors]): a list of colors, where each color
855
+ corresponds to each mask or box in the image. Refer to 'matplotlib.colors'
856
+ for full list of formats that the colors are accepted in.
857
+ Returns:
858
+ output (VisImage): image object with visualizations.
859
+ """
860
+ num_instances = len(boxes)
861
+
862
+ if assigned_colors is None:
863
+ # assigned_colors = [random_color(rgb=True, maximum=1) for _ in range(num_instances)]
864
+ assigned_colors = [instance_color(rgb=True, idx=i, maximum=1) for i in range(num_instances)]
865
+ if num_instances == 0:
866
+ return self.output
867
+
868
+ # Display in largest to smallest order to reduce occlusion.
869
+ if boxes is not None:
870
+ areas = boxes[:, 2] * boxes[:, 3]
871
+
872
+ sorted_idxs = np.argsort(-areas).tolist()
873
+ # Re-order overlapped instances in descending order.
874
+ boxes = boxes[sorted_idxs]
875
+ labels = [labels[k] for k in sorted_idxs] if labels is not None else None
876
+ colors = [assigned_colors[idx] for idx in sorted_idxs]
877
+
878
+ for i in range(num_instances):
879
+ self.draw_rotated_box_with_label(
880
+ boxes[i], edge_color=colors[i], label=labels[i] if labels is not None else None
881
+ )
882
+
883
+ return self.output
884
+
885
+ def draw_and_connect_keypoints(self, keypoints):
886
+ """
887
+ Draws keypoints of an instance and follows the rules for keypoint connections
888
+ to draw lines between appropriate keypoints. This follows color heuristics for
889
+ line color.
890
+ Args:
891
+ keypoints (Tensor): a tensor of shape (K, 3), where K is the number of keypoints
892
+ and the last dimension corresponds to (x, y, probability).
893
+ Returns:
894
+ output (VisImage): image object with visualizations.
895
+ """
896
+ visible = {}
897
+ keypoint_names = self.metadata.get("keypoint_names")
898
+ for idx, keypoint in enumerate(keypoints):
899
+
900
+ # draw keypoint
901
+ x, y, prob = keypoint
902
+ if prob > self.keypoint_threshold:
903
+ self.draw_circle((x, y), color=_RED)
904
+ if keypoint_names:
905
+ keypoint_name = keypoint_names[idx]
906
+ visible[keypoint_name] = (x, y)
907
+
908
+ if self.metadata.get("keypoint_connection_rules"):
909
+ for kp0, kp1, color in self.metadata.keypoint_connection_rules:
910
+ if kp0 in visible and kp1 in visible:
911
+ x0, y0 = visible[kp0]
912
+ x1, y1 = visible[kp1]
913
+ color = tuple(x / 255.0 for x in color)
914
+ self.draw_line([x0, x1], [y0, y1], color=color)
915
+
916
+ # draw lines from nose to mid-shoulder and mid-shoulder to mid-hip
917
+ # Note that this strategy is specific to person keypoints.
918
+ # For other keypoints, it should just do nothing
919
+ try:
920
+ ls_x, ls_y = visible["left_shoulder"]
921
+ rs_x, rs_y = visible["right_shoulder"]
922
+ mid_shoulder_x, mid_shoulder_y = (ls_x + rs_x) / 2, (ls_y + rs_y) / 2
923
+ except KeyError:
924
+ pass
925
+ else:
926
+ # draw line from nose to mid-shoulder
927
+ nose_x, nose_y = visible.get("nose", (None, None))
928
+ if nose_x is not None:
929
+ self.draw_line([nose_x, mid_shoulder_x], [nose_y, mid_shoulder_y], color=_RED)
930
+
931
+ try:
932
+ # draw line from mid-shoulder to mid-hip
933
+ lh_x, lh_y = visible["left_hip"]
934
+ rh_x, rh_y = visible["right_hip"]
935
+ except KeyError:
936
+ pass
937
+ else:
938
+ mid_hip_x, mid_hip_y = (lh_x + rh_x) / 2, (lh_y + rh_y) / 2
939
+ self.draw_line([mid_hip_x, mid_shoulder_x], [mid_hip_y, mid_shoulder_y], color=_RED)
940
+ return self.output
941
+
942
+ """
943
+ Primitive drawing functions:
944
+ """
945
+
946
+ def draw_text(
947
+ self,
948
+ text,
949
+ position,
950
+ *,
951
+ font_size=None,
952
+ color="g",
953
+ horizontal_alignment="center",
954
+ rotation=0,
955
+ ):
956
+ """
957
+ Args:
958
+ text (str): class label
959
+ position (tuple): a tuple of the x and y coordinates to place text on image.
960
+ font_size (int, optional): font of the text. If not provided, a font size
961
+ proportional to the image width is calculated and used.
962
+ color: color of the text. Refer to `matplotlib.colors` for full list
963
+ of formats that are accepted.
964
+ horizontal_alignment (str): see `matplotlib.text.Text`
965
+ rotation: rotation angle in degrees CCW
966
+ Returns:
967
+ output (VisImage): image object with text drawn.
968
+ """
969
+ if not font_size:
970
+ font_size = self._default_font_size
971
+
972
+ # since the text background is dark, we don't want the text to be dark
973
+ color = np.maximum(list(mplc.to_rgb(color)), 0.2)
974
+ color[np.argmax(color)] = max(0.8, np.max(color))
975
+
976
+ x, y = position
977
+ self.output.ax.text(
978
+ x,
979
+ y,
980
+ text,
981
+ size=font_size * self.output.scale,
982
+ family="sans-serif",
983
+ bbox={"facecolor": "black", "alpha": 0.8, "pad": 0.7, "edgecolor": "none"},
984
+ verticalalignment="top",
985
+ horizontalalignment=horizontal_alignment,
986
+ color=color,
987
+ zorder=10,
988
+ rotation=rotation,
989
+ )
990
+ return self.output
991
+
992
+ def draw_box(self, box_coord, alpha=1.0, edge_color="g", line_style="-"):
993
+ """
994
+ Args:
995
+ box_coord (tuple): a tuple containing x0, y0, x1, y1 coordinates, where x0 and y0
996
+ are the coordinates of the image's top left corner. x1 and y1 are the
997
+ coordinates of the image's bottom right corner.
998
+ alpha (float): blending efficient. Smaller values lead to more transparent masks.
999
+ edge_color: color of the outline of the box. Refer to `matplotlib.colors`
1000
+ for full list of formats that are accepted.
1001
+ line_style (string): the string to use to create the outline of the boxes.
1002
+ Returns:
1003
+ output (VisImage): image object with box drawn.
1004
+ """
1005
+ x0, y0, x1, y1 = box_coord
1006
+ width = x1 - x0
1007
+ height = y1 - y0
1008
+
1009
+ linewidth = 2
1010
+
1011
+ self.output.ax.add_patch(
1012
+ mpl.patches.Rectangle(
1013
+ (x0, y0),
1014
+ width,
1015
+ height,
1016
+ fill=False,
1017
+ edgecolor=edge_color,
1018
+ linewidth=linewidth * self.output.scale,
1019
+ alpha=alpha,
1020
+ linestyle=line_style,
1021
+ )
1022
+ )
1023
+ return self.output
1024
+
1025
+ def draw_rotated_box_with_label(
1026
+ self, rotated_box, alpha=0.5, edge_color="g", line_style="-", label=None
1027
+ ):
1028
+ """
1029
+ Draw a rotated box with label on its top-left corner.
1030
+ Args:
1031
+ rotated_box (tuple): a tuple containing (cnt_x, cnt_y, w, h, angle),
1032
+ where cnt_x and cnt_y are the center coordinates of the box.
1033
+ w and h are the width and height of the box. angle represents how
1034
+ many degrees the box is rotated CCW with regard to the 0-degree box.
1035
+ alpha (float): blending efficient. Smaller values lead to more transparent masks.
1036
+ edge_color: color of the outline of the box. Refer to `matplotlib.colors`
1037
+ for full list of formats that are accepted.
1038
+ line_style (string): the string to use to create the outline of the boxes.
1039
+ label (string): label for rotated box. It will not be rendered when set to None.
1040
+ Returns:
1041
+ output (VisImage): image object with box drawn.
1042
+ """
1043
+ cnt_x, cnt_y, w, h, angle = rotated_box
1044
+ area = w * h
1045
+ # use thinner lines when the box is small
1046
+ linewidth = self._default_font_size / (
1047
+ 6 if area < _SMALL_OBJECT_AREA_THRESH * self.output.scale else 3
1048
+ )
1049
+
1050
+ theta = angle * math.pi / 180.0
1051
+ c = math.cos(theta)
1052
+ s = math.sin(theta)
1053
+ rect = [(-w / 2, h / 2), (-w / 2, -h / 2), (w / 2, -h / 2), (w / 2, h / 2)]
1054
+ # x: left->right ; y: top->down
1055
+ rotated_rect = [(s * yy + c * xx + cnt_x, c * yy - s * xx + cnt_y) for (xx, yy) in rect]
1056
+ for k in range(4):
1057
+ j = (k + 1) % 4
1058
+ self.draw_line(
1059
+ [rotated_rect[k][0], rotated_rect[j][0]],
1060
+ [rotated_rect[k][1], rotated_rect[j][1]],
1061
+ color=edge_color,
1062
+ linestyle="--" if k == 1 else line_style,
1063
+ linewidth=linewidth,
1064
+ )
1065
+
1066
+ if label is not None:
1067
+ text_pos = rotated_rect[1] # topleft corner
1068
+
1069
+ height_ratio = h / np.sqrt(self.output.height * self.output.width)
1070
+ label_color = self._change_color_brightness(edge_color, brightness_factor=0.7)
1071
+ font_size = (
1072
+ np.clip((height_ratio - 0.02) / 0.08 + 1, 1.2, 2) * 0.5 * self._default_font_size
1073
+ )
1074
+ self.draw_text(label, text_pos, color=label_color, font_size=font_size, rotation=angle)
1075
+
1076
+ return self.output
1077
+
1078
+ def draw_circle(self, circle_coord, color, radius=3):
1079
+ """
1080
+ Args:
1081
+ circle_coord (list(int) or tuple(int)): contains the x and y coordinates
1082
+ of the center of the circle.
1083
+ color: color of the polygon. Refer to `matplotlib.colors` for a full list of
1084
+ formats that are accepted.
1085
+ radius (int): radius of the circle.
1086
+ Returns:
1087
+ output (VisImage): image object with box drawn.
1088
+ """
1089
+ x, y = circle_coord
1090
+ self.output.ax.add_patch(
1091
+ mpl.patches.Circle(circle_coord, radius=radius, fill=True, color=color)
1092
+ )
1093
+ return self.output
1094
+
1095
+ def draw_line(self, x_data, y_data, color, linestyle="-", linewidth=None):
1096
+ """
1097
+ Args:
1098
+ x_data (list[int]): a list containing x values of all the points being drawn.
1099
+ Length of list should match the length of y_data.
1100
+ y_data (list[int]): a list containing y values of all the points being drawn.
1101
+ Length of list should match the length of x_data.
1102
+ color: color of the line. Refer to `matplotlib.colors` for a full list of
1103
+ formats that are accepted.
1104
+ linestyle: style of the line. Refer to `matplotlib.lines.Line2D`
1105
+ for a full list of formats that are accepted.
1106
+ linewidth (float or None): width of the line. When it's None,
1107
+ a default value will be computed and used.
1108
+ Returns:
1109
+ output (VisImage): image object with line drawn.
1110
+ """
1111
+ if linewidth is None:
1112
+ linewidth = self._default_font_size / 3
1113
+ linewidth = max(linewidth, 1)
1114
+ self.output.ax.add_line(
1115
+ mpl.lines.Line2D(
1116
+ x_data,
1117
+ y_data,
1118
+ linewidth=linewidth * self.output.scale,
1119
+ color=color,
1120
+ linestyle=linestyle,
1121
+ )
1122
+ )
1123
+ return self.output
1124
+
1125
+ def draw_binary_mask(
1126
+ self, binary_mask, color=None, *, edge_color=None, text=None, alpha=0.5, area_threshold=10, is_text=True,
1127
+ ):
1128
+ """
1129
+ Args:
1130
+ binary_mask (ndarray): numpy array of shape (H, W), where H is the image height and
1131
+ W is the image width. Each value in the array is either a 0 or 1 value of uint8
1132
+ type.
1133
+ color: color of the mask. Refer to `matplotlib.colors` for a full list of
1134
+ formats that are accepted. If None, will pick a random color.
1135
+ edge_color: color of the polygon edges. Refer to `matplotlib.colors` for a
1136
+ full list of formats that are accepted.
1137
+ text (str): if None, will be drawn on the object
1138
+ alpha (float): blending efficient. Smaller values lead to more transparent masks.
1139
+ area_threshold (float): a connected component smaller than this area will not be shown.
1140
+ Returns:
1141
+ output (VisImage): image object with mask drawn.
1142
+ """
1143
+ if color is None:
1144
+ color = random_color(rgb=True, maximum=1)
1145
+ color = mplc.to_rgb(color)
1146
+
1147
+ has_valid_segment = False
1148
+ binary_mask = binary_mask.astype("uint8") # opencv needs uint8
1149
+ mask = GenericMask(binary_mask, self.output.height, self.output.width)
1150
+ shape2d = (binary_mask.shape[0], binary_mask.shape[1])
1151
+
1152
+ if not mask.has_holes:
1153
+ # draw polygons for regular masks
1154
+ for segment in mask.polygons:
1155
+ area = mask_util.area(mask_util.frPyObjects([segment], shape2d[0], shape2d[1]))
1156
+ if area < (area_threshold or 0):
1157
+ continue
1158
+ has_valid_segment = True
1159
+ segment = segment.reshape(-1, 2)
1160
+ self.draw_polygon(segment, color=color, edge_color=edge_color, alpha=alpha)
1161
+ else:
1162
+ # TODO: Use Path/PathPatch to draw vector graphics:
1163
+ # https://stackoverflow.com/questions/8919719/how-to-plot-a-complex-polygon
1164
+ rgba = np.zeros(shape2d + (4,), dtype="float32")
1165
+ rgba[:, :, :3] = color
1166
+ rgba[:, :, 3] = (mask.mask == 1).astype("float32") * alpha
1167
+ has_valid_segment = True
1168
+ self.output.ax.imshow(rgba, extent=(0, self.output.width, self.output.height, 0))
1169
+
1170
+ if is_text:
1171
+ if text is not None and has_valid_segment:
1172
+ lighter_color = self._change_color_brightness(color, brightness_factor=0.7)
1173
+ self._draw_text_in_mask(binary_mask, text, lighter_color)
1174
+ return self.output
1175
+
1176
+ def draw_soft_mask(self, soft_mask, color=None, *, text=None, alpha=0.5):
1177
+ """
1178
+ Args:
1179
+ soft_mask (ndarray): float array of shape (H, W), each value in [0, 1].
1180
+ color: color of the mask. Refer to `matplotlib.colors` for a full list of
1181
+ formats that are accepted. If None, will pick a random color.
1182
+ text (str): if None, will be drawn on the object
1183
+ alpha (float): blending efficient. Smaller values lead to more transparent masks.
1184
+ Returns:
1185
+ output (VisImage): image object with mask drawn.
1186
+ """
1187
+ if color is None:
1188
+ color = random_color(rgb=True, maximum=1)
1189
+ color = mplc.to_rgb(color)
1190
+
1191
+ shape2d = (soft_mask.shape[0], soft_mask.shape[1])
1192
+ rgba = np.zeros(shape2d + (4,), dtype="float32")
1193
+ rgba[:, :, :3] = color
1194
+ rgba[:, :, 3] = soft_mask * alpha
1195
+ self.output.ax.imshow(rgba, extent=(0, self.output.width, self.output.height, 0))
1196
+
1197
+ if text is not None:
1198
+ lighter_color = self._change_color_brightness(color, brightness_factor=0.7)
1199
+ binary_mask = (soft_mask > 0.5).astype("uint8")
1200
+ # self._draw_text_in_mask(binary_mask, text, lighter_color)
1201
+ return self.output
1202
+
1203
+ def draw_polygon(self, segment, color, edge_color=None, alpha=0.5):
1204
+ """
1205
+ Args:
1206
+ segment: numpy array of shape Nx2, containing all the points in the polygon.
1207
+ color: color of the polygon. Refer to `matplotlib.colors` for a full list of
1208
+ formats that are accepted.
1209
+ edge_color: color of the polygon edges. Refer to `matplotlib.colors` for a
1210
+ full list of formats that are accepted. If not provided, a darker shade
1211
+ of the polygon color will be used instead.
1212
+ alpha (float): blending efficient. Smaller values lead to more transparent masks.
1213
+ Returns:
1214
+ output (VisImage): image object with polygon drawn.
1215
+ """
1216
+ if edge_color is None:
1217
+ # make edge color darker than the polygon color
1218
+ if alpha > 0.8:
1219
+ edge_color = self._change_color_brightness(color, brightness_factor=-0.7)
1220
+ else:
1221
+ edge_color = color
1222
+ edge_color = mplc.to_rgb(edge_color) + (1,)
1223
+
1224
+ polygon = mpl.patches.Polygon(
1225
+ segment,
1226
+ fill=True,
1227
+ facecolor=mplc.to_rgb(color) + (alpha,),
1228
+ edgecolor=edge_color,
1229
+ linewidth=max(self._default_font_size // 15 * self.output.scale, 1),
1230
+ )
1231
+ self.output.ax.add_patch(polygon)
1232
+ return self.output
1233
+
1234
+ """
1235
+ Internal methods:
1236
+ """
1237
+
1238
+ def _jitter(self, color):
1239
+ """
1240
+ Randomly modifies given color to produce a slightly different color than the color given.
1241
+ Args:
1242
+ color (tuple[double]): a tuple of 3 elements, containing the RGB values of the color
1243
+ picked. The values in the list are in the [0.0, 1.0] range.
1244
+ Returns:
1245
+ jittered_color (tuple[double]): a tuple of 3 elements, containing the RGB values of the
1246
+ color after being jittered. The values in the list are in the [0.0, 1.0] range.
1247
+ """
1248
+ color = mplc.to_rgb(color)
1249
+ vec = np.random.rand(3)
1250
+ # better to do it in another color space
1251
+ vec = vec / np.linalg.norm(vec) * 0.5
1252
+ res = np.clip(vec + color, 0, 1)
1253
+ return tuple(res)
1254
+
1255
+ def _create_grayscale_image(self, mask=None):
1256
+ """
1257
+ Create a grayscale version of the original image.
1258
+ The colors in masked area, if given, will be kept.
1259
+ """
1260
+ img_bw = self.img.astype("f4").mean(axis=2)
1261
+ img_bw = np.stack([img_bw] * 3, axis=2)
1262
+ if mask is not None:
1263
+ img_bw[mask] = self.img[mask]
1264
+ return img_bw
1265
+
1266
+ def _change_color_brightness(self, color, brightness_factor):
1267
+ """
1268
+ Depending on the brightness_factor, gives a lighter or darker color i.e. a color with
1269
+ less or more saturation than the original color.
1270
+ Args:
1271
+ color: color of the polygon. Refer to `matplotlib.colors` for a full list of
1272
+ formats that are accepted.
1273
+ brightness_factor (float): a value in [-1.0, 1.0] range. A lightness factor of
1274
+ 0 will correspond to no change, a factor in [-1.0, 0) range will result in
1275
+ a darker color and a factor in (0, 1.0] range will result in a lighter color.
1276
+ Returns:
1277
+ modified_color (tuple[double]): a tuple containing the RGB values of the
1278
+ modified color. Each value in the tuple is in the [0.0, 1.0] range.
1279
+ """
1280
+ assert brightness_factor >= -1.0 and brightness_factor <= 1.0
1281
+ color = mplc.to_rgb(color)
1282
+ polygon_color = colorsys.rgb_to_hls(*mplc.to_rgb(color))
1283
+ modified_lightness = polygon_color[1] + (brightness_factor * polygon_color[1])
1284
+ modified_lightness = 0.0 if modified_lightness < 0.0 else modified_lightness
1285
+ modified_lightness = 1.0 if modified_lightness > 1.0 else modified_lightness
1286
+ modified_color = colorsys.hls_to_rgb(polygon_color[0], modified_lightness, polygon_color[2])
1287
+ return modified_color
1288
+
1289
+ def _convert_boxes(self, boxes):
1290
+ """
1291
+ Convert different format of boxes to an NxB array, where B = 4 or 5 is the box dimension.
1292
+ """
1293
+ if isinstance(boxes, Boxes) or isinstance(boxes, RotatedBoxes):
1294
+ return boxes.tensor.detach().numpy()
1295
+ else:
1296
+ return np.asarray(boxes)
1297
+
1298
+ def _convert_masks(self, masks_or_polygons):
1299
+ """
1300
+ Convert different format of masks or polygons to a tuple of masks and polygons.
1301
+ Returns:
1302
+ list[GenericMask]:
1303
+ """
1304
+
1305
+ m = masks_or_polygons
1306
+ if isinstance(m, PolygonMasks):
1307
+ m = m.polygons
1308
+ if isinstance(m, BitMasks):
1309
+ m = m.tensor.numpy()
1310
+ if isinstance(m, torch.Tensor):
1311
+ m = m.numpy()
1312
+ ret = []
1313
+ for x in m:
1314
+ if isinstance(x, GenericMask):
1315
+ ret.append(x)
1316
+ else:
1317
+ ret.append(GenericMask(x, self.output.height, self.output.width))
1318
+ return ret
1319
+
1320
+ def _draw_text_in_mask(self, binary_mask, text, color):
1321
+ """
1322
+ Find proper places to draw text given a binary mask.
1323
+ """
1324
+ # TODO sometimes drawn on wrong objects. the heuristics here can improve.
1325
+ _num_cc, cc_labels, stats, centroids = cv2.connectedComponentsWithStats(binary_mask, 8)
1326
+ if stats[1:, -1].size == 0:
1327
+ return
1328
+ largest_component_id = np.argmax(stats[1:, -1]) + 1
1329
+
1330
+ # draw text on the largest component, as well as other very large components.
1331
+ for cid in range(1, _num_cc):
1332
+ if cid == largest_component_id or stats[cid, -1] > _LARGE_MASK_AREA_THRESH:
1333
+ # median is more stable than centroid
1334
+ # center = centroids[largest_component_id]
1335
+ center = np.median((cc_labels == cid).nonzero(), axis=1)[::-1]
1336
+ self.draw_text(text, center, color=color)
1337
+
1338
+ def _convert_keypoints(self, keypoints):
1339
+ if isinstance(keypoints, Keypoints):
1340
+ keypoints = keypoints.tensor
1341
+ keypoints = np.asarray(keypoints)
1342
+ return keypoints
1343
+
1344
+ def get_output(self):
1345
+ """
1346
+ Returns:
1347
+ output (VisImage): the image output containing the visualizations added
1348
+ to the image.
1349
+ """
1350
+ return self.output
examples/ade20k.jpeg ADDED

Git LFS Details

  • SHA256: 99f7af15a7bd66f3d2ad8b98f1b41941c27407a35d303f9b70724cb231a36098
  • Pointer size: 130 Bytes
  • Size of remote file: 52.7 kB
examples/cityscapes.png ADDED

Git LFS Details

  • SHA256: 55d6e21e5358e9df0fde8a5fb0b50cf1aac1b5865dc454ebb23774d66ca6be2a
  • Pointer size: 132 Bytes
  • Size of remote file: 2.3 MB
examples/coco.jpeg ADDED

Git LFS Details

  • SHA256: 83981537a7baeafbeb9c8cb67b3484dc26433f574b3685d021fa537e277e4726
  • Pointer size: 131 Bytes
  • Size of remote file: 134 kB
gradio_app.py ADDED
@@ -0,0 +1,194 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import subprocess
2
+ import torch
3
+ if torch.cuda.is_available():
4
+ subprocess.run('sh deform_setup.sh', shell=True)
5
+
6
+ print("Installed the dependencies!")
7
+
8
+ import numpy as np
9
+ from PIL import Image
10
+ import cv2
11
+ import imutils
12
+
13
+ from detectron2.config import get_cfg
14
+ from detectron2.projects.deeplab import add_deeplab_config
15
+ from detectron2.data import MetadataCatalog
16
+
17
+ from oneformer import (
18
+ add_oneformer_config,
19
+ add_common_config,
20
+ add_swin_config,
21
+ add_dinat_config,
22
+ )
23
+
24
+ from demo.defaults import DefaultPredictor
25
+ from demo.visualizer import Visualizer, ColorMode
26
+
27
+ import gradio as gr
28
+
29
+ KEY_DICT = {"Cityscapes (19 classes)": "cityscapes",
30
+ "COCO (133 classes)": "coco",
31
+ "ADE20K (150 classes)": "ade20k",}
32
+
33
+ SWIN_CFG_DICT = {"cityscapes": "configs/cityscapes/oneformer_swin_large_IN21k_384_bs16_90k.yaml",
34
+ "coco": "configs/coco/oneformer_swin_large_IN21k_384_bs16_100ep.yaml",
35
+ "ade20k": "configs/ade20k/oneformer_swin_large_IN21k_384_bs16_160k.yaml",}
36
+
37
+ SWIN_MODEL_DICT = {"cityscapes": "models/250_16_swin_l_oneformer_cityscapes_90k.pth",
38
+ "coco": "models/150_16_swin_l_oneformer_coco_100ep.pth",
39
+ "ade20k": "models/250_16_swin_l_oneformer_ade20k_160k.pth"}
40
+
41
+ DINAT_CFG_DICT = {"cityscapes": "configs/cityscapes/oneformer_dinat_large_bs16_90k.yaml",
42
+ "coco": "configs/coco/oneformer_dinat_large_bs16_100ep.yaml",
43
+ "ade20k": "configs/ade20k/oneformer_dinat_large_IN21k_384_bs16_160k.yaml",}
44
+
45
+ DINAT_MODEL_DICT = {"cityscapes": "models/250_16_dinat_l_oneformer_cityscapes_90k.pth",
46
+ "coco": "models/150_16_dinat_l_oneformer_coco_100ep.pth",
47
+ "ade20k": "models/250_16_dinat_l_oneformer_ade20k_160k.pth"}
48
+
49
+ MODEL_DICT = {"DiNAT-L": DINAT_MODEL_DICT,
50
+ "Swin-L": SWIN_MODEL_DICT }
51
+
52
+ CFG_DICT = {"DiNAT-L": DINAT_CFG_DICT,
53
+ "Swin-L": SWIN_CFG_DICT }
54
+
55
+ WIDTH_DICT = {"cityscapes": 512,
56
+ "coco": 512,
57
+ "ade20k": 640}
58
+
59
+ cpu_device = torch.device("cpu")
60
+
61
+ PREDICTORS = {
62
+ "DiNAT-L": {
63
+ "Cityscapes (19 classes)": None,
64
+ "COCO (133 classes)": None,
65
+ "ADE20K (150 classes)": None
66
+ },
67
+ "Swin-L": {
68
+ "Cityscapes (19 classes)": None,
69
+ "COCO (133 classes)": None,
70
+ "ADE20K (150 classes)": None
71
+ }
72
+ }
73
+
74
+ def setup_predictors():
75
+ for dataset in ["Cityscapes (19 classes)", "COCO (133 classes)", "ADE20K (150 classes)"]:
76
+ for backbone in ["DiNAT-L", "Swin-L"]:
77
+ cfg = setup_cfg(dataset, backbone)
78
+ PREDICTORS[backbone][dataset] = DefaultPredictor(cfg)
79
+
80
+ def setup_cfg(dataset, backbone):
81
+ # load config from file and command-line arguments
82
+ cfg = get_cfg()
83
+ add_deeplab_config(cfg)
84
+ add_common_config(cfg)
85
+ add_swin_config(cfg)
86
+ add_oneformer_config(cfg)
87
+ add_dinat_config(cfg)
88
+ dataset = KEY_DICT[dataset]
89
+ cfg_path = CFG_DICT[backbone][dataset]
90
+ cfg.merge_from_file(cfg_path)
91
+ if torch.cuda.is_available():
92
+ cfg.MODEL.DEVICE = 'cuda'
93
+ else:
94
+ cfg.MODEL.DEVICE = 'cpu'
95
+ cfg.MODEL.WEIGHTS = MODEL_DICT[backbone][dataset]
96
+ cfg.freeze()
97
+ return cfg
98
+
99
+ def setup_modules(dataset, backbone):
100
+ cfg = setup_cfg(dataset, backbone)
101
+ # predictor = DefaultPredictor(cfg)
102
+ predictor = PREDICTORS[backbone][dataset]
103
+ metadata = MetadataCatalog.get(
104
+ cfg.DATASETS.TEST_PANOPTIC[0] if len(cfg.DATASETS.TEST_PANOPTIC) else "__unused"
105
+ )
106
+ if 'cityscapes_fine_sem_seg_val' in cfg.DATASETS.TEST_PANOPTIC[0]:
107
+ from cityscapesscripts.helpers.labels import labels
108
+ stuff_colors = [k.color for k in labels if k.trainId != 255]
109
+ metadata = metadata.set(stuff_colors=stuff_colors)
110
+
111
+ return predictor, metadata
112
+
113
+ def panoptic_run(img, predictor, metadata):
114
+ visualizer = Visualizer(img[:, :, ::-1], metadata=metadata, instance_mode=ColorMode.IMAGE)
115
+ predictions = predictor(img, "panoptic")
116
+ panoptic_seg, segments_info = predictions["panoptic_seg"]
117
+ out = visualizer.draw_panoptic_seg_predictions(
118
+ panoptic_seg.to(cpu_device), segments_info, alpha=0.5
119
+ )
120
+ visualizer_map = Visualizer(img[:, :, ::-1], is_img=False, metadata=metadata, instance_mode=ColorMode.IMAGE)
121
+ out_map = visualizer_map.draw_panoptic_seg_predictions(
122
+ panoptic_seg.to(cpu_device), segments_info, alpha=1, is_text=False
123
+ )
124
+ return out, out_map
125
+
126
+ def instance_run(img, predictor, metadata):
127
+ visualizer = Visualizer(img[:, :, ::-1], metadata=metadata, instance_mode=ColorMode.IMAGE)
128
+ predictions = predictor(img, "instance")
129
+ instances = predictions["instances"].to(cpu_device)
130
+ out = visualizer.draw_instance_predictions(predictions=instances, alpha=0.5)
131
+ visualizer_map = Visualizer(img[:, :, ::-1], is_img=False, metadata=metadata, instance_mode=ColorMode.IMAGE)
132
+ out_map = visualizer_map.draw_instance_predictions(predictions=instances, alpha=1, is_text=False)
133
+ return out, out_map
134
+
135
+ def semantic_run(img, predictor, metadata):
136
+ visualizer = Visualizer(img[:, :, ::-1], metadata=metadata, instance_mode=ColorMode.IMAGE)
137
+ predictions = predictor(img, "semantic")
138
+ out = visualizer.draw_sem_seg(
139
+ predictions["sem_seg"].argmax(dim=0).to(cpu_device), alpha=0.5
140
+ )
141
+ visualizer_map = Visualizer(img[:, :, ::-1], is_img=False, metadata=metadata, instance_mode=ColorMode.IMAGE)
142
+ out_map = visualizer_map.draw_sem_seg(
143
+ predictions["sem_seg"].argmax(dim=0).to(cpu_device), alpha=1, is_text=False
144
+ )
145
+ return out, out_map
146
+
147
+ TASK_INFER = {"the task is panoptic": panoptic_run, "the task is instance": instance_run, "the task is semantic": semantic_run}
148
+
149
+ def segment(path, task, dataset, backbone):
150
+ predictor, metadata = setup_modules(dataset, backbone)
151
+ img = cv2.imread(path)
152
+ width = WIDTH_DICT[KEY_DICT[dataset]]
153
+ img = imutils.resize(img, width=width)
154
+ out, out_map = TASK_INFER[task](img, predictor, metadata)
155
+ out = Image.fromarray(out.get_image())
156
+ out_map = Image.fromarray(out_map.get_image())
157
+ return out, out_map
158
+
159
+ title = "OneFormer: One Transformer to Rule Universal Image Segmentation"
160
+
161
+ description = "<p style='color: #E0B941; font-size: 16px; font-weight: w600; text-align: center'> <a style='color: #E0B941;' href='https://praeclarumjj3.github.io/oneformer/' target='_blank'>Project Page</a> | <a style='color: #E0B941;' href='https://arxiv.org/abs/2211.06220' target='_blank'>OneFormer: One Transformer to Rule Universal Image Segmentation</a> | <a style='color: #E0B941;' href='https://github.com/SHI-Labs/OneFormer' target='_blank'>Github</a></p>" \
162
+ + "<p style='color:royalblue; margin: 10px; font-size: 16px; font-weight: w400;'> \
163
+ [Note: Inference on CPU may take upto 2 minutes.] This is the official gradio demo for our paper <span style='color:#E0B941;'>OneFormer: One Transformer to Rule Universal Image Segmentation</span> To use OneFormer: <br> \
164
+ (1) <span style='color:#E0B941;'>Upload an Image</span> or <span style='color:#E0B941;'> select a sample image from the examples</span> <br> \
165
+ (2) Select the value of the <span style='color:#E0B941;'>Task Token Input</span> <br>\
166
+ (3) Select the <span style='color:#E0B941;'>Model</span> </p>"
167
+
168
+ # article =
169
+
170
+ # css = ".image-preview {height: 32rem; width: auto;} .output-image {height: 32rem; width: auto;} .panel-buttons { display: flex; flex-direction: row;}"
171
+
172
+ setup_predictors()
173
+
174
+ gradio_inputs = [gr.Image(source="upload", tool=None, label="Input Image",type="filepath"),
175
+ gr.inputs.Radio(choices=["the task is panoptic" ,"the task is instance", "the task is semantic"], type="value", default="the task is panoptic", label="Task Token Input"),
176
+ gr.inputs.Radio(choices=["COCO (133 classes)" ,"Cityscapes (19 classes)", "ADE20K (150 classes)"], type="value", default="Cityscapes (19 classes)", label="Model"),
177
+ gr.inputs.Radio(choices=["DiNAT-L" ,"Swin-L"], type="value", default="DiNAT-L", label="Backbone"),
178
+ ]
179
+ gradio_outputs = [gr.Image(type="pil", label="Segmentation Overlay"), gr.Image(type="pil", label="Segmentation Map")]
180
+
181
+
182
+ examples = [["examples/coco.jpeg", "the task is panoptic", "COCO (133 classes)", "DiNAT-L"],
183
+ ["examples/cityscapes.png", "the task is panoptic", "Cityscapes (19 classes)", "DiNAT-L"],
184
+ ["examples/ade20k.jpeg", "the task is panoptic", "ADE20K (150 classes)", "DiNAT-L"]]
185
+
186
+
187
+ iface = gr.Interface(fn=segment, inputs=gradio_inputs,
188
+ outputs=gradio_outputs,
189
+ examples_per_page=5,
190
+ allow_flagging="never",
191
+ examples=examples, title=title,
192
+ description=description)
193
+
194
+ iface.launch(enable_queue=True)
oneformer/.DS_Store ADDED
Binary file (6.15 kB). View file
 
oneformer/__init__.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ from . import data # register all new datasets
3
+ from . import modeling
4
+
5
+ # config
6
+ from .config import *
7
+
8
+ # models
9
+ from .oneformer_model import OneFormer
oneformer/config.py ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ # Copyright (c) Facebook, Inc. and its affiliates.
3
+ from detectron2.config import CfgNode as CN
4
+
5
+ __all__ = ["add_common_config", "add_oneformer_config", "add_swin_config",
6
+ "add_dinat_config", "add_beit_adapter_config", "add_convnext_config"]
7
+
8
+ def add_common_config(cfg):
9
+ """
10
+ Add config for common configuration
11
+ """
12
+ # data config
13
+ # select the dataset mapper
14
+ cfg.INPUT.DATASET_MAPPER_NAME = "oneformer_unified"
15
+ # Color augmentation
16
+ cfg.INPUT.COLOR_AUG_SSD = False
17
+ # We retry random cropping until no single category in semantic segmentation GT occupies more
18
+ # than `SINGLE_CATEGORY_MAX_AREA` part of the crop.
19
+ cfg.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA = 1.0
20
+ # Pad image and segmentation GT in dataset mapper.
21
+ cfg.INPUT.SIZE_DIVISIBILITY = -1
22
+
23
+ cfg.INPUT.TASK_SEQ_LEN = 77
24
+ cfg.INPUT.MAX_SEQ_LEN = 77
25
+
26
+ cfg.INPUT.TASK_PROB = CN()
27
+ cfg.INPUT.TASK_PROB.SEMANTIC = 0.33
28
+ cfg.INPUT.TASK_PROB.INSTANCE = 0.66
29
+
30
+ # test dataset
31
+ cfg.DATASETS.TEST_PANOPTIC = ("",)
32
+ cfg.DATASETS.TEST_INSTANCE = ("",)
33
+ cfg.DATASETS.TEST_SEMANTIC = ("",)
34
+
35
+ # solver config
36
+ # weight decay on embedding
37
+ cfg.SOLVER.WEIGHT_DECAY_EMBED = 0.0
38
+ # optimizer
39
+ cfg.SOLVER.OPTIMIZER = "ADAMW"
40
+ cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1
41
+
42
+ # wandb
43
+ cfg.WANDB = CN()
44
+ cfg.WANDB.PROJECT = "unified_dense_recognition"
45
+ cfg.WANDB.NAME = None
46
+
47
+ cfg.MODEL.IS_TRAIN = False
48
+ cfg.MODEL.IS_DEMO = True
49
+
50
+ # text encoder config
51
+ cfg.MODEL.TEXT_ENCODER = CN()
52
+
53
+ cfg.MODEL.TEXT_ENCODER.WIDTH = 256
54
+ cfg.MODEL.TEXT_ENCODER.CONTEXT_LENGTH = 77
55
+ cfg.MODEL.TEXT_ENCODER.NUM_LAYERS = 12
56
+ cfg.MODEL.TEXT_ENCODER.VOCAB_SIZE = 49408
57
+ cfg.MODEL.TEXT_ENCODER.PROJ_NUM_LAYERS = 2
58
+ cfg.MODEL.TEXT_ENCODER.N_CTX = 16
59
+
60
+ # mask_former inference config
61
+ cfg.MODEL.TEST = CN()
62
+ cfg.MODEL.TEST.SEMANTIC_ON = True
63
+ cfg.MODEL.TEST.INSTANCE_ON = False
64
+ cfg.MODEL.TEST.PANOPTIC_ON = False
65
+ cfg.MODEL.TEST.DETECTION_ON = False
66
+ cfg.MODEL.TEST.OBJECT_MASK_THRESHOLD = 0.0
67
+ cfg.MODEL.TEST.OVERLAP_THRESHOLD = 0.0
68
+ cfg.MODEL.TEST.SEM_SEG_POSTPROCESSING_BEFORE_INFERENCE = False
69
+ cfg.MODEL.TEST.TASK = "panoptic"
70
+
71
+ # TEST AUG Slide
72
+ cfg.TEST.AUG.IS_SLIDE = False
73
+ cfg.TEST.AUG.CROP_SIZE = (640, 640)
74
+ cfg.TEST.AUG.STRIDE = (426, 426)
75
+ cfg.TEST.AUG.SCALE = (2048, 640)
76
+ cfg.TEST.AUG.SETR_MULTI_SCALE = True
77
+ cfg.TEST.AUG.KEEP_RATIO = True
78
+ cfg.TEST.AUG.SIZE_DIVISOR = 32
79
+
80
+ # pixel decoder config
81
+ cfg.MODEL.SEM_SEG_HEAD.MASK_DIM = 256
82
+ # adding transformer in pixel decoder
83
+ cfg.MODEL.SEM_SEG_HEAD.TRANSFORMER_ENC_LAYERS = 0
84
+ # pixel decoder
85
+ cfg.MODEL.SEM_SEG_HEAD.PIXEL_DECODER_NAME = "BasePixelDecoder"
86
+ cfg.MODEL.SEM_SEG_HEAD.SEM_EMBED_DIM = 256
87
+ cfg.MODEL.SEM_SEG_HEAD.INST_EMBED_DIM = 256
88
+
89
+ # LSJ aug
90
+ cfg.INPUT.IMAGE_SIZE = 1024
91
+ cfg.INPUT.MIN_SCALE = 0.1
92
+ cfg.INPUT.MAX_SCALE = 2.0
93
+
94
+ # MSDeformAttn encoder configs
95
+ cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES = ["res3", "res4", "res5"]
96
+ cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_POINTS = 4
97
+ cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_HEADS = 8
98
+
99
+ def add_oneformer_config(cfg):
100
+ """
101
+ Add config for ONE_FORMER.
102
+ """
103
+
104
+ # mask_former model config
105
+ cfg.MODEL.ONE_FORMER = CN()
106
+
107
+ # loss
108
+ cfg.MODEL.ONE_FORMER.DEEP_SUPERVISION = True
109
+ cfg.MODEL.ONE_FORMER.NO_OBJECT_WEIGHT = 0.1
110
+ cfg.MODEL.ONE_FORMER.CLASS_WEIGHT = 1.0
111
+ cfg.MODEL.ONE_FORMER.DICE_WEIGHT = 1.0
112
+ cfg.MODEL.ONE_FORMER.MASK_WEIGHT = 20.0
113
+ cfg.MODEL.ONE_FORMER.CONTRASTIVE_WEIGHT = 0.5
114
+ cfg.MODEL.ONE_FORMER.CONTRASTIVE_TEMPERATURE = 0.07
115
+
116
+ # transformer config
117
+ cfg.MODEL.ONE_FORMER.NHEADS = 8
118
+ cfg.MODEL.ONE_FORMER.DROPOUT = 0.1
119
+ cfg.MODEL.ONE_FORMER.DIM_FEEDFORWARD = 2048
120
+ cfg.MODEL.ONE_FORMER.ENC_LAYERS = 0
121
+ cfg.MODEL.ONE_FORMER.CLASS_DEC_LAYERS = 2
122
+ cfg.MODEL.ONE_FORMER.DEC_LAYERS = 6
123
+ cfg.MODEL.ONE_FORMER.PRE_NORM = False
124
+
125
+ cfg.MODEL.ONE_FORMER.HIDDEN_DIM = 256
126
+ cfg.MODEL.ONE_FORMER.NUM_OBJECT_QUERIES = 120
127
+ cfg.MODEL.ONE_FORMER.NUM_OBJECT_CTX = 16
128
+ cfg.MODEL.ONE_FORMER.USE_TASK_NORM = True
129
+
130
+ cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE = "res5"
131
+ cfg.MODEL.ONE_FORMER.ENFORCE_INPUT_PROJ = False
132
+
133
+ # Sometimes `backbone.size_divisibility` is set to 0 for some backbone (e.g. ResNet)
134
+ # you can use this config to override
135
+ cfg.MODEL.ONE_FORMER.SIZE_DIVISIBILITY = 32
136
+
137
+ # transformer module
138
+ cfg.MODEL.ONE_FORMER.TRANSFORMER_DECODER_NAME = "ContrastiveMultiScaleMaskedTransformerDecoder"
139
+
140
+ # point loss configs
141
+ # Number of points sampled during training for a mask point head.
142
+ cfg.MODEL.ONE_FORMER.TRAIN_NUM_POINTS = 112 * 112
143
+ # Oversampling parameter for PointRend point sampling during training. Parameter `k` in the
144
+ # original paper.
145
+ cfg.MODEL.ONE_FORMER.OVERSAMPLE_RATIO = 3.0
146
+ # Importance sampling parameter for PointRend point sampling during training. Parametr `beta` in
147
+ # the original paper.
148
+ cfg.MODEL.ONE_FORMER.IMPORTANCE_SAMPLE_RATIO = 0.75
149
+
150
+ def add_swin_config(cfg):
151
+ """
152
+ Add config forSWIN Backbone.
153
+ """
154
+
155
+ # swin transformer backbone
156
+ cfg.MODEL.SWIN = CN()
157
+ cfg.MODEL.SWIN.PRETRAIN_IMG_SIZE = 224
158
+ cfg.MODEL.SWIN.PATCH_SIZE = 4
159
+ cfg.MODEL.SWIN.EMBED_DIM = 96
160
+ cfg.MODEL.SWIN.DEPTHS = [2, 2, 6, 2]
161
+ cfg.MODEL.SWIN.NUM_HEADS = [3, 6, 12, 24]
162
+ cfg.MODEL.SWIN.WINDOW_SIZE = 7
163
+ cfg.MODEL.SWIN.MLP_RATIO = 4.0
164
+ cfg.MODEL.SWIN.QKV_BIAS = True
165
+ cfg.MODEL.SWIN.QK_SCALE = None
166
+ cfg.MODEL.SWIN.DROP_RATE = 0.0
167
+ cfg.MODEL.SWIN.ATTN_DROP_RATE = 0.0
168
+ cfg.MODEL.SWIN.DROP_PATH_RATE = 0.3
169
+ cfg.MODEL.SWIN.APE = False
170
+ cfg.MODEL.SWIN.PATCH_NORM = True
171
+ cfg.MODEL.SWIN.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
172
+ cfg.MODEL.SWIN.USE_CHECKPOINT = False
173
+ ## Semask additions
174
+ cfg.MODEL.SWIN.SEM_WINDOW_SIZE = 7
175
+ cfg.MODEL.SWIN.NUM_SEM_BLOCKS = 1
176
+
177
+ def add_dinat_config(cfg):
178
+ """
179
+ Add config for NAT Backbone.
180
+ """
181
+
182
+ # DINAT transformer backbone
183
+ cfg.MODEL.DiNAT = CN()
184
+ cfg.MODEL.DiNAT.DEPTHS = [3, 4, 18, 5]
185
+ cfg.MODEL.DiNAT.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
186
+ cfg.MODEL.DiNAT.EMBED_DIM = 64
187
+ cfg.MODEL.DiNAT.MLP_RATIO = 3.0
188
+ cfg.MODEL.DiNAT.NUM_HEADS = [2, 4, 8, 16]
189
+ cfg.MODEL.DiNAT.DROP_PATH_RATE = 0.2
190
+ cfg.MODEL.DiNAT.KERNEL_SIZE = 7
191
+ cfg.MODEL.DiNAT.DILATIONS = [[1, 16, 1], [1, 4, 1, 8], [1, 2, 1, 3, 1, 4], [1, 2, 1, 2, 1]]
192
+ cfg.MODEL.DiNAT.OUT_INDICES = (0, 1, 2, 3)
193
+ cfg.MODEL.DiNAT.QKV_BIAS = True
194
+ cfg.MODEL.DiNAT.QK_SCALE = None
195
+ cfg.MODEL.DiNAT.DROP_RATE = 0
196
+ cfg.MODEL.DiNAT.ATTN_DROP_RATE = 0.
197
+ cfg.MODEL.DiNAT.IN_PATCH_SIZE = 4
198
+
199
+ def add_convnext_config(cfg):
200
+ """
201
+ Add config for ConvNeXt Backbone.
202
+ """
203
+
204
+ # swin transformer backbone
205
+ cfg.MODEL.CONVNEXT = CN()
206
+ cfg.MODEL.CONVNEXT.IN_CHANNELS = 3
207
+ cfg.MODEL.CONVNEXT.DEPTHS = [3, 3, 27, 3]
208
+ cfg.MODEL.CONVNEXT.DIMS = [192, 384, 768, 1536]
209
+ cfg.MODEL.CONVNEXT.DROP_PATH_RATE = 0.4
210
+ cfg.MODEL.CONVNEXT.LSIT = 1.0
211
+ cfg.MODEL.CONVNEXT.OUT_INDICES = [0, 1, 2, 3]
212
+ cfg.MODEL.CONVNEXT.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
213
+
214
+ def add_beit_adapter_config(cfg):
215
+ """
216
+ Add config for BEiT Adapter Backbone.
217
+ """
218
+
219
+ # beit adapter backbone
220
+ cfg.MODEL.BEiTAdapter = CN()
221
+ cfg.MODEL.BEiTAdapter.IMG_SIZE = 640
222
+ cfg.MODEL.BEiTAdapter.PATCH_SIZE = 16
223
+ cfg.MODEL.BEiTAdapter.EMBED_DIM = 1024
224
+ cfg.MODEL.BEiTAdapter.DEPTH = 24
225
+ cfg.MODEL.BEiTAdapter.NUM_HEADS = 16
226
+ cfg.MODEL.BEiTAdapter.MLP_RATIO = 4
227
+ cfg.MODEL.BEiTAdapter.QKV_BIAS = True
228
+ cfg.MODEL.BEiTAdapter.USE_ABS_POS_EMB = False
229
+ cfg.MODEL.BEiTAdapter.USE_REL_POS_BIAS = True
230
+ cfg.MODEL.BEiTAdapter.INIT_VALUES = 1e-6
231
+ cfg.MODEL.BEiTAdapter.DROP_PATH_RATE = 0.3
232
+ cfg.MODEL.BEiTAdapter.CONV_INPLANE = 64
233
+ cfg.MODEL.BEiTAdapter.N_POINTS = 4
234
+ cfg.MODEL.BEiTAdapter.DEFORM_NUM_HEADS = 16
235
+ cfg.MODEL.BEiTAdapter.CFFN_RATIO = 0.25
236
+ cfg.MODEL.BEiTAdapter.DEFORM_RATIO = 0.5
237
+ cfg.MODEL.BEiTAdapter.WITH_CP = True
238
+ cfg.MODEL.BEiTAdapter.INTERACTION_INDEXES=[[0, 5], [6, 11], [12, 17], [18, 23]]
239
+ cfg.MODEL.BEiTAdapter.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
oneformer/data/__init__.py ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ from . import datasets
oneformer/data/bpe_simple_vocab_16e6.txt ADDED
The diff for this file is too large to render. See raw diff
 
oneformer/data/bpe_simple_vocab_16e6.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
3
+ size 1356917
oneformer/data/build.py ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ from typing import Any, Callable, Dict, List, Optional, Union
3
+ import torch.utils.data as torchdata
4
+
5
+ from detectron2.config import configurable
6
+
7
+
8
+ from detectron2.data.common import DatasetFromList, MapDataset
9
+ from detectron2.data.dataset_mapper import DatasetMapper
10
+ from detectron2.data.samplers import (
11
+ InferenceSampler,
12
+ )
13
+ from detectron2.data.build import (
14
+ get_detection_dataset_dicts,
15
+ trivial_batch_collator
16
+ )
17
+ """
18
+ This file contains the default logic to build a dataloader for training or testing.
19
+ """
20
+
21
+ __all__ = [
22
+ "build_detection_test_loader",
23
+ ]
24
+
25
+
26
+ def _test_loader_from_config(cfg, dataset_name, mapper=None):
27
+ """
28
+ Uses the given `dataset_name` argument (instead of the names in cfg), because the
29
+ standard practice is to evaluate each test set individually (not combining them).
30
+ """
31
+ if isinstance(dataset_name, str):
32
+ dataset_name = [dataset_name]
33
+
34
+ dataset = get_detection_dataset_dicts(
35
+ dataset_name,
36
+ filter_empty=False,
37
+ proposal_files=[
38
+ cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(x)] for x in dataset_name
39
+ ]
40
+ if cfg.MODEL.LOAD_PROPOSALS
41
+ else None,
42
+ )
43
+ if mapper is None:
44
+ mapper = DatasetMapper(cfg, False)
45
+ return {
46
+ "dataset": dataset,
47
+ "mapper": mapper,
48
+ "num_workers": cfg.DATALOADER.NUM_WORKERS,
49
+ "sampler": InferenceSampler(len(dataset))
50
+ if not isinstance(dataset, torchdata.IterableDataset)
51
+ else None,
52
+ }
53
+
54
+
55
+ @configurable(from_config=_test_loader_from_config)
56
+ def build_detection_test_loader(
57
+ dataset: Union[List[Any], torchdata.Dataset],
58
+ *,
59
+ mapper: Callable[[Dict[str, Any]], Any],
60
+ sampler: Optional[torchdata.Sampler] = None,
61
+ batch_size: int = 1,
62
+ num_workers: int = 0,
63
+ collate_fn: Optional[Callable[[List[Any]], Any]] = None,
64
+ ) -> torchdata.DataLoader:
65
+ """
66
+ Similar to `build_detection_train_loader`, with default batch size = 1,
67
+ and sampler = :class:`InferenceSampler`. This sampler coordinates all workers
68
+ to produce the exact set of all samples.
69
+
70
+ Args:
71
+ dataset: a list of dataset dicts,
72
+ or a pytorch dataset (either map-style or iterable). They can be obtained
73
+ by using :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`.
74
+ mapper: a callable which takes a sample (dict) from dataset
75
+ and returns the format to be consumed by the model.
76
+ When using cfg, the default choice is ``DatasetMapper(cfg, is_train=False)``.
77
+ sampler: a sampler that produces
78
+ indices to be applied on ``dataset``. Default to :class:`InferenceSampler`,
79
+ which splits the dataset across all workers. Sampler must be None
80
+ if `dataset` is iterable.
81
+ batch_size: the batch size of the data loader to be created.
82
+ Default to 1 image per worker since this is the standard when reporting
83
+ inference time in papers.
84
+ num_workers: number of parallel data loading workers
85
+ collate_fn: same as the argument of `torch.utils.data.DataLoader`.
86
+ Defaults to do no collation and return a list of data.
87
+
88
+ Returns:
89
+ DataLoader: a torch DataLoader, that loads the given detection
90
+ dataset, with test-time transformation and batching.
91
+
92
+ Examples:
93
+ ::
94
+ data_loader = build_detection_test_loader(
95
+ DatasetRegistry.get("my_test"),
96
+ mapper=DatasetMapper(...))
97
+
98
+ # or, instantiate with a CfgNode:
99
+ data_loader = build_detection_test_loader(cfg, "my_test")
100
+ """
101
+ if isinstance(dataset, list):
102
+ dataset = DatasetFromList(dataset, copy=False)
103
+ if mapper is not None:
104
+ dataset = MapDataset(dataset, mapper)
105
+ if isinstance(dataset, torchdata.IterableDataset):
106
+ assert sampler is None, "sampler must be None if dataset is IterableDataset"
107
+ else:
108
+ if sampler is None:
109
+ sampler = InferenceSampler(len(dataset))
110
+ return torchdata.DataLoader(
111
+ dataset,
112
+ batch_size=batch_size,
113
+ sampler=sampler,
114
+ drop_last=False,
115
+ num_workers=num_workers,
116
+ collate_fn=trivial_batch_collator if collate_fn is None else collate_fn,
117
+ )
oneformer/data/dataset_mappers/__init__.py ADDED
@@ -0,0 +1 @@
 
 
1
+
oneformer/data/dataset_mappers/coco_unified_new_baseline_dataset_mapper.py ADDED
@@ -0,0 +1,341 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/data/dataset_mappers/coco_panoptic_new_baseline_dataset_mapper.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import copy
7
+ import logging
8
+
9
+ import numpy as np
10
+ import torch
11
+
12
+ from detectron2.data import MetadataCatalog
13
+ from detectron2.config import configurable
14
+ from detectron2.data import detection_utils as utils
15
+ from detectron2.data import transforms as T
16
+ from detectron2.structures import BitMasks, Instances
17
+ from oneformer.utils.box_ops import masks_to_boxes
18
+ from oneformer.data.tokenizer import SimpleTokenizer, Tokenize
19
+
20
+ __all__ = ["COCOUnifiedNewBaselineDatasetMapper"]
21
+
22
+
23
+ def build_transform_gen(cfg, is_train):
24
+ """
25
+ Create a list of default :class:`Augmentation` from config.
26
+ Now it includes resizing and flipping.
27
+ Returns:
28
+ list[Augmentation]
29
+ """
30
+ assert is_train, "Only support training augmentation"
31
+ image_size = cfg.INPUT.IMAGE_SIZE
32
+ min_scale = cfg.INPUT.MIN_SCALE
33
+ max_scale = cfg.INPUT.MAX_SCALE
34
+
35
+ augmentation = []
36
+
37
+ if cfg.INPUT.RANDOM_FLIP != "none":
38
+ augmentation.append(
39
+ T.RandomFlip(
40
+ horizontal=cfg.INPUT.RANDOM_FLIP == "horizontal",
41
+ vertical=cfg.INPUT.RANDOM_FLIP == "vertical",
42
+ )
43
+ )
44
+
45
+ augmentation.extend([
46
+ T.ResizeScale(
47
+ min_scale=min_scale, max_scale=max_scale, target_height=image_size, target_width=image_size
48
+ ),
49
+ T.FixedSizeCrop(crop_size=(image_size, image_size)),
50
+ ])
51
+
52
+ return augmentation
53
+
54
+
55
+ # This is specifically designed for the COCO dataset.
56
+ class COCOUnifiedNewBaselineDatasetMapper:
57
+ """
58
+ A callable which takes a dataset dict in Detectron2 Dataset format,
59
+ and map it into a format used by OneFormer.
60
+
61
+ This dataset mapper applies the same transformation as DETR for COCO panoptic segmentation.
62
+
63
+ The callable currently does the following:
64
+
65
+ 1. Read the image from "file_name"
66
+ 2. Applies geometric transforms to the image and annotation
67
+ 3. Find and applies suitable cropping to the image and annotation
68
+ 4. Prepare image and annotation to Tensors
69
+ """
70
+
71
+ @configurable
72
+ def __init__(
73
+ self,
74
+ is_train=True,
75
+ *,
76
+ num_queries,
77
+ tfm_gens,
78
+ meta,
79
+ image_format,
80
+ max_seq_len,
81
+ task_seq_len,
82
+ semantic_prob,
83
+ instance_prob,
84
+ ):
85
+ """
86
+ NOTE: this interface is experimental.
87
+ Args:
88
+ is_train: for training or inference
89
+ augmentations: a list of augmentations or deterministic transforms to apply
90
+ crop_gen: crop augmentation
91
+ tfm_gens: data augmentation
92
+ image_format: an image format supported by :func:`detection_utils.read_image`.
93
+ """
94
+ self.tfm_gens = tfm_gens
95
+ logging.getLogger(__name__).info(
96
+ "[COCOUnifiedNewBaselineDatasetMapper] Full TransformGens used in training: {}".format(
97
+ str(self.tfm_gens)
98
+ )
99
+ )
100
+
101
+ self.img_format = image_format
102
+ self.is_train = is_train
103
+ self.meta = meta
104
+ self.ignore_label = self.meta.ignore_label
105
+ self.num_queries = num_queries
106
+
107
+ self.things = []
108
+ for k,v in self.meta.thing_dataset_id_to_contiguous_id.items():
109
+ self.things.append(v)
110
+ self.class_names = self.meta.stuff_classes
111
+ self.text_tokenizer = Tokenize(SimpleTokenizer(), max_seq_len=max_seq_len)
112
+ self.task_tokenizer = Tokenize(SimpleTokenizer(), max_seq_len=task_seq_len)
113
+ self.semantic_prob = semantic_prob
114
+ self.instance_prob = instance_prob
115
+
116
+ @classmethod
117
+ def from_config(cls, cfg, is_train=True):
118
+ # Build augmentation
119
+ tfm_gens = build_transform_gen(cfg, is_train)
120
+ dataset_names = cfg.DATASETS.TRAIN
121
+ meta = MetadataCatalog.get(dataset_names[0])
122
+
123
+ ret = {
124
+ "is_train": is_train,
125
+ "meta": meta,
126
+ "tfm_gens": tfm_gens,
127
+ "image_format": cfg.INPUT.FORMAT,
128
+ "num_queries": cfg.MODEL.ONE_FORMER.NUM_OBJECT_QUERIES - cfg.MODEL.TEXT_ENCODER.N_CTX,
129
+ "task_seq_len": cfg.INPUT.TASK_SEQ_LEN,
130
+ "max_seq_len": cfg.INPUT.MAX_SEQ_LEN,
131
+ "semantic_prob": cfg.INPUT.TASK_PROB.SEMANTIC,
132
+ "instance_prob": cfg.INPUT.TASK_PROB.INSTANCE,
133
+ }
134
+ return ret
135
+
136
+ def _get_semantic_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
137
+ instances = Instances(image_shape)
138
+
139
+ classes = []
140
+ texts = ["a semantic photo"] * self.num_queries
141
+ masks = []
142
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
143
+
144
+ for segment_info in segments_info:
145
+ class_id = segment_info["category_id"]
146
+ if not segment_info["iscrowd"]:
147
+ mask = pan_seg_gt == segment_info["id"]
148
+ if not np.all(mask == False):
149
+ if class_id not in classes:
150
+ cls_name = self.class_names[class_id]
151
+ classes.append(class_id)
152
+ masks.append(mask)
153
+ num_class_obj[cls_name] += 1
154
+ else:
155
+ idx = classes.index(class_id)
156
+ masks[idx] += mask
157
+ masks[idx] = np.clip(masks[idx], 0, 1).astype(np.bool)
158
+ label[mask] = class_id
159
+
160
+ num = 0
161
+ for i, cls_name in enumerate(self.class_names):
162
+ if num_class_obj[cls_name] > 0:
163
+ for _ in range(num_class_obj[cls_name]):
164
+ if num >= len(texts):
165
+ break
166
+ texts[num] = f"a photo with a {cls_name}"
167
+ num += 1
168
+
169
+ classes = np.array(classes)
170
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
171
+ if len(masks) == 0:
172
+ # Some image does not have annotation (all ignored)
173
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
174
+ instances.gt_bboxes = torch.zeros((0, 4))
175
+ else:
176
+ masks = BitMasks(
177
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
178
+ )
179
+ instances.gt_masks = masks.tensor
180
+ # Placeholder bounding boxes for stuff regions. Note that these are not used during training.
181
+ instances.gt_bboxes = torch.stack([torch.tensor([0., 0., 1., 1.])] * instances.gt_masks.shape[0])
182
+ return instances, texts, label
183
+
184
+ def _get_instance_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
185
+ instances = Instances(image_shape)
186
+
187
+ classes = []
188
+ texts = ["an instance photo"] * self.num_queries
189
+ masks = []
190
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
191
+
192
+ for segment_info in segments_info:
193
+ class_id = segment_info["category_id"]
194
+ if class_id in self.things:
195
+ if not segment_info["iscrowd"]:
196
+ mask = pan_seg_gt == segment_info["id"]
197
+ if not np.all(mask == False):
198
+ cls_name = self.class_names[class_id]
199
+ classes.append(class_id)
200
+ masks.append(mask)
201
+ num_class_obj[cls_name] += 1
202
+ label[mask] = class_id
203
+
204
+ num = 0
205
+ for i, cls_name in enumerate(self.class_names):
206
+ if num_class_obj[cls_name] > 0:
207
+ for _ in range(num_class_obj[cls_name]):
208
+ if num >= len(texts):
209
+ break
210
+ texts[num] = f"a photo with a {cls_name}"
211
+ num += 1
212
+
213
+ classes = np.array(classes)
214
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
215
+ if len(masks) == 0:
216
+ # Some image does not have annotation (all ignored)
217
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
218
+ instances.gt_bboxes = torch.zeros((0, 4))
219
+ else:
220
+ masks = BitMasks(
221
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
222
+ )
223
+ instances.gt_masks = masks.tensor
224
+ instances.gt_bboxes = masks_to_boxes(instances.gt_masks)
225
+ return instances, texts, label
226
+
227
+ def _get_panoptic_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
228
+ instances = Instances(image_shape)
229
+
230
+ classes = []
231
+ texts = ["a panoptic photo"] * self.num_queries
232
+ masks = []
233
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
234
+
235
+ for segment_info in segments_info:
236
+ class_id = segment_info["category_id"]
237
+ if not segment_info["iscrowd"]:
238
+ mask = pan_seg_gt == segment_info["id"]
239
+ if not np.all(mask == False):
240
+ cls_name = self.class_names[class_id]
241
+ classes.append(class_id)
242
+ masks.append(mask)
243
+ num_class_obj[cls_name] += 1
244
+ label[mask] = class_id
245
+
246
+ num = 0
247
+ for i, cls_name in enumerate(self.class_names):
248
+ if num_class_obj[cls_name] > 0:
249
+ for _ in range(num_class_obj[cls_name]):
250
+ if num >= len(texts):
251
+ break
252
+ texts[num] = f"a photo with a {cls_name}"
253
+ num += 1
254
+
255
+ classes = np.array(classes)
256
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
257
+ if len(masks) == 0:
258
+ # Some image does not have annotation (all ignored)
259
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
260
+ instances.gt_bboxes = torch.zeros((0, 4))
261
+ else:
262
+ masks = BitMasks(
263
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
264
+ )
265
+ instances.gt_masks = masks.tensor
266
+ instances.gt_bboxes = masks_to_boxes(instances.gt_masks)
267
+ for i in range(instances.gt_classes.shape[0]):
268
+ # Placeholder bounding boxes for stuff regions. Note that these are not used during training.
269
+ if instances.gt_classes[i].item() not in self.things:
270
+ instances.gt_bboxes[i] = torch.tensor([0., 0., 1., 1.])
271
+ return instances, texts, label
272
+
273
+ def __call__(self, dataset_dict):
274
+ """
275
+ Args:
276
+ dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
277
+
278
+ Returns:
279
+ dict: a format that builtin models in detectron2 accept
280
+ """
281
+ dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below
282
+ image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
283
+ utils.check_image_size(dataset_dict, image)
284
+
285
+ image, transforms = T.apply_transform_gens(self.tfm_gens, image)
286
+ image_shape = image.shape[:2] # h, w
287
+
288
+ # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
289
+ # but not efficient on large generic data structures due to the use of pickle & mp.Queue.
290
+ # Therefore it's important to use torch.Tensor.
291
+ dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
292
+
293
+ if not self.is_train:
294
+ # USER: Modify this if you want to keep them for some reason.
295
+ dataset_dict.pop("annotations", None)
296
+ return dataset_dict
297
+
298
+ # semantic segmentation
299
+ if "sem_seg_file_name" in dataset_dict:
300
+ # PyTorch transformation not implemented for uint16, so converting it to double first
301
+ sem_seg_gt = utils.read_image(dataset_dict.pop("sem_seg_file_name")).astype("double")
302
+ sem_seg_gt = transforms.apply_segmentation(sem_seg_gt)
303
+ else:
304
+ sem_seg_gt = None
305
+
306
+ if "pan_seg_file_name" in dataset_dict:
307
+ pan_seg_gt = utils.read_image(dataset_dict.pop("pan_seg_file_name"), "RGB")
308
+ segments_info = dataset_dict["segments_info"]
309
+
310
+ # apply the same transformation to panoptic segmentation
311
+ pan_seg_gt = transforms.apply_segmentation(pan_seg_gt)
312
+
313
+ from panopticapi.utils import rgb2id
314
+ pan_seg_gt = rgb2id(pan_seg_gt)
315
+
316
+ prob_task = np.random.uniform(0,1.)
317
+
318
+ num_class_obj = {}
319
+
320
+ for name in self.class_names:
321
+ num_class_obj[name] = 0
322
+
323
+ if prob_task < self.semantic_prob:
324
+ task = "The task is semantic"
325
+ instances, text, sem_seg = self._get_semantic_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
326
+ elif prob_task < self.instance_prob:
327
+ task = "The task is instance"
328
+ instances, text, sem_seg = self._get_instance_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
329
+ else:
330
+ task = "The task is panoptic"
331
+ instances, text, sem_seg = self._get_panoptic_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
332
+
333
+
334
+ dataset_dict["sem_seg"] = torch.from_numpy(sem_seg).long()
335
+ dataset_dict["instances"] = instances
336
+ dataset_dict["orig_shape"] = image_shape
337
+ dataset_dict["task"] = task
338
+ dataset_dict["text"] = text
339
+ dataset_dict["thing_ids"] = self.things
340
+
341
+ return dataset_dict
oneformer/data/dataset_mappers/dataset_mapper.py ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/dataset_mapper.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import copy
7
+ import logging
8
+ import numpy as np
9
+ from typing import List, Optional, Union
10
+ import torch
11
+
12
+ from detectron2.config import configurable
13
+
14
+ from detectron2.data import detection_utils as utils
15
+ from detectron2.data import transforms as T
16
+ from oneformer.data.tokenizer import SimpleTokenizer, Tokenize
17
+
18
+ __all__ = ["DatasetMapper"]
19
+
20
+
21
+ class DatasetMapper:
22
+ """
23
+ A callable which takes a dataset dict in Detectron2 Dataset format,
24
+ and map it into a format used by the model.
25
+
26
+ This is the default callable to be used to map your dataset dict into training data.
27
+ You may need to follow it to implement your own one for customized logic,
28
+ such as a different way to read or transform images.
29
+ See :doc:`/tutorials/data_loading` for details.
30
+
31
+ The callable currently does the following:
32
+
33
+ 1. Read the image from "file_name"
34
+ 2. Applies cropping/geometric transforms to the image and annotations
35
+ 3. Prepare data and annotations to Tensor and :class:`Instances`
36
+ """
37
+
38
+ @configurable
39
+ def __init__(
40
+ self,
41
+ is_train: bool,
42
+ *,
43
+ augmentations: List[Union[T.Augmentation, T.Transform]],
44
+ image_format: str,
45
+ task_seq_len: int,
46
+ task: str = "panoptic",
47
+ use_instance_mask: bool = False,
48
+ use_keypoint: bool = False,
49
+ instance_mask_format: str = "polygon",
50
+ keypoint_hflip_indices: Optional[np.ndarray] = None,
51
+ precomputed_proposal_topk: Optional[int] = None,
52
+ recompute_boxes: bool = False,
53
+ ):
54
+ """
55
+ NOTE: this interface is experimental.
56
+
57
+ Args:
58
+ is_train: whether it's used in training or inference
59
+ augmentations: a list of augmentations or deterministic transforms to apply
60
+ image_format: an image format supported by :func:`detection_utils.read_image`.
61
+ use_instance_mask: whether to process instance segmentation annotations, if available
62
+ use_keypoint: whether to process keypoint annotations if available
63
+ instance_mask_format: one of "polygon" or "bitmask". Process instance segmentation
64
+ masks into this format.
65
+ keypoint_hflip_indices: see :func:`detection_utils.create_keypoint_hflip_indices`
66
+ precomputed_proposal_topk: if given, will load pre-computed
67
+ proposals from dataset_dict and keep the top k proposals for each image.
68
+ recompute_boxes: whether to overwrite bounding box annotations
69
+ by computing tight bounding boxes from instance mask annotations.
70
+ """
71
+ if recompute_boxes:
72
+ assert use_instance_mask, "recompute_boxes requires instance masks"
73
+ # fmt: off
74
+ self.is_train = is_train
75
+ self.augmentations = T.AugmentationList(augmentations)
76
+ self.image_format = image_format
77
+ self.use_instance_mask = use_instance_mask
78
+ self.instance_mask_format = instance_mask_format
79
+ self.use_keypoint = use_keypoint
80
+ self.keypoint_hflip_indices = keypoint_hflip_indices
81
+ self.proposal_topk = precomputed_proposal_topk
82
+ self.recompute_boxes = recompute_boxes
83
+ self.task_tokenizer = Tokenize(SimpleTokenizer(), max_seq_len=task_seq_len)
84
+ self.task = task
85
+ assert self.task in ["panoptic", "semantic", "instance"]
86
+
87
+ # fmt: on
88
+ logger = logging.getLogger(__name__)
89
+ mode = "training" if is_train else "inference"
90
+ logger.info(f"[DatasetMapper] Augmentations used in {mode}: {augmentations}")
91
+
92
+ @classmethod
93
+ def from_config(cls, cfg, is_train: bool = True):
94
+ augs = utils.build_augmentation(cfg, is_train)
95
+ if cfg.INPUT.CROP.ENABLED and is_train:
96
+ augs.insert(0, T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE))
97
+ recompute_boxes = cfg.MODEL.MASK_ON
98
+ else:
99
+ recompute_boxes = False
100
+
101
+ ret = {
102
+ "is_train": is_train,
103
+ "augmentations": augs,
104
+ "image_format": cfg.INPUT.FORMAT,
105
+ "use_instance_mask": cfg.MODEL.MASK_ON,
106
+ "instance_mask_format": cfg.INPUT.MASK_FORMAT,
107
+ "use_keypoint": cfg.MODEL.KEYPOINT_ON,
108
+ "task_seq_len": cfg.INPUT.TASK_SEQ_LEN,
109
+ "recompute_boxes": recompute_boxes,
110
+ "task": cfg.MODEL.TEST.TASK,
111
+ }
112
+
113
+ if cfg.MODEL.KEYPOINT_ON:
114
+ ret["keypoint_hflip_indices"] = utils.create_keypoint_hflip_indices(cfg.DATASETS.TRAIN)
115
+
116
+ if cfg.MODEL.LOAD_PROPOSALS:
117
+ ret["precomputed_proposal_topk"] = (
118
+ cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN
119
+ if is_train
120
+ else cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST
121
+ )
122
+ return ret
123
+
124
+ def _transform_annotations(self, dataset_dict, transforms, image_shape):
125
+ # USER: Modify this if you want to keep them for some reason.
126
+ for anno in dataset_dict["annotations"]:
127
+ if not self.use_instance_mask:
128
+ anno.pop("segmentation", None)
129
+ if not self.use_keypoint:
130
+ anno.pop("keypoints", None)
131
+
132
+ # USER: Implement additional transformations if you have other types of data
133
+ annos = [
134
+ utils.transform_instance_annotations(
135
+ obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices
136
+ )
137
+ for obj in dataset_dict.pop("annotations")
138
+ if obj.get("iscrowd", 0) == 0
139
+ ]
140
+ instances = utils.annotations_to_instances(
141
+ annos, image_shape, mask_format=self.instance_mask_format
142
+ )
143
+
144
+ # After transforms such as cropping are applied, the bounding box may no longer
145
+ # tightly bound the object. As an example, imagine a triangle object
146
+ # [(0,0), (2,0), (0,2)] cropped by a box [(1,0),(2,2)] (XYXY format). The tight
147
+ # bounding box of the cropped triangle should be [(1,0),(2,1)], which is not equal to
148
+ # the intersection of original bounding box and the cropping box.
149
+ if self.recompute_boxes:
150
+ instances.gt_boxes = instances.gt_masks.get_bounding_boxes()
151
+ dataset_dict["instances"] = utils.filter_empty_instances(instances)
152
+
153
+ def __call__(self, dataset_dict):
154
+ """
155
+ Args:
156
+ dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
157
+
158
+ Returns:
159
+ dict: a format that builtin models in detectron2 accept
160
+ """
161
+ dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below
162
+ # USER: Write your own image loading if it's not from a file
163
+ image = utils.read_image(dataset_dict["file_name"], format=self.image_format)
164
+ utils.check_image_size(dataset_dict, image)
165
+
166
+ task = f"The task is {self.task}"
167
+ dataset_dict["task"] = task
168
+
169
+ # USER: Remove if you don't do semantic/panoptic segmentation.
170
+ if "sem_seg_file_name" in dataset_dict:
171
+ sem_seg_gt = utils.read_image(dataset_dict.pop("sem_seg_file_name"), "L").squeeze(2)
172
+ else:
173
+ sem_seg_gt = None
174
+
175
+ aug_input = T.AugInput(image, sem_seg=sem_seg_gt)
176
+ transforms = self.augmentations(aug_input)
177
+ image, sem_seg_gt = aug_input.image, aug_input.sem_seg
178
+
179
+ image_shape = image.shape[:2] # h, w
180
+ # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
181
+ # but not efficient on large generic data structures due to the use of pickle & mp.Queue.
182
+ # Therefore it's important to use torch.Tensor.
183
+ dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
184
+ if sem_seg_gt is not None:
185
+ dataset_dict["sem_seg"] = torch.as_tensor(sem_seg_gt.astype("long"))
186
+
187
+ # USER: Remove if you don't use pre-computed proposals.
188
+ # Most users would not need this feature.
189
+ if self.proposal_topk is not None:
190
+ utils.transform_proposals(
191
+ dataset_dict, image_shape, transforms, proposal_topk=self.proposal_topk
192
+ )
193
+
194
+ if not self.is_train:
195
+ # USER: Modify this if you want to keep them for some reason.
196
+ dataset_dict.pop("annotations", None)
197
+ dataset_dict.pop("sem_seg_file_name", None)
198
+ return dataset_dict
199
+
200
+ if "annotations" in dataset_dict:
201
+ self._transform_annotations(dataset_dict, transforms, image_shape)
202
+
203
+ return dataset_dict
oneformer/data/dataset_mappers/oneformer_unified_dataset_mapper.py ADDED
@@ -0,0 +1,375 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/data/dataset_mappers/mask_former_panoptic_dataset_mapper.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import copy
7
+ import logging
8
+ import os
9
+
10
+ import numpy as np
11
+ import torch
12
+ from torch.nn import functional as F
13
+
14
+ from detectron2.config import configurable
15
+ from detectron2.data import detection_utils as utils
16
+ from detectron2.data import transforms as T
17
+ from detectron2.structures import BitMasks, Instances
18
+ from detectron2.data import MetadataCatalog
19
+ from detectron2.projects.point_rend import ColorAugSSDTransform
20
+ from oneformer.utils.box_ops import masks_to_boxes
21
+ from oneformer.data.tokenizer import SimpleTokenizer, Tokenize
22
+
23
+ __all__ = ["OneFormerUnifiedDatasetMapper"]
24
+
25
+
26
+ class OneFormerUnifiedDatasetMapper:
27
+ """
28
+ A callable which takes a dataset dict in Detectron2 Dataset format,
29
+ and map it into a format used by OneFormer for universal segmentation.
30
+
31
+ The callable currently does the following:
32
+
33
+ 1. Read the image from "file_name"
34
+ 2. Applies geometric transforms to the image and annotation
35
+ 3. Find and applies suitable cropping to the image and annotation
36
+ 4. Prepare image and annotation to Tensors
37
+ """
38
+
39
+ @configurable
40
+ def __init__(
41
+ self,
42
+ is_train=True,
43
+ *,
44
+ name,
45
+ num_queries,
46
+ meta,
47
+ augmentations,
48
+ image_format,
49
+ ignore_label,
50
+ size_divisibility,
51
+ task_seq_len,
52
+ max_seq_len,
53
+ semantic_prob,
54
+ instance_prob,
55
+ ):
56
+ """
57
+ NOTE: this interface is experimental.
58
+ Args:
59
+ is_train: for training or inference
60
+ augmentations: a list of augmentations or deterministic transforms to apply
61
+ image_format: an image format supported by :func:`detection_utils.read_image`.
62
+ ignore_label: the label that is ignored to evaluation
63
+ size_divisibility: pad image size to be divisible by this value
64
+ """
65
+ self.is_train = is_train
66
+ self.meta = meta
67
+ self.name = name
68
+ self.tfm_gens = augmentations
69
+ self.img_format = image_format
70
+ self.ignore_label = ignore_label
71
+ self.size_divisibility = size_divisibility
72
+ self.num_queries = num_queries
73
+
74
+ logger = logging.getLogger(__name__)
75
+ mode = "training" if is_train else "inference"
76
+ logger.info(f"[{self.__class__.__name__}] Augmentations used in {mode}: {augmentations}")
77
+
78
+ self.things = []
79
+ for k,v in self.meta.thing_dataset_id_to_contiguous_id.items():
80
+ self.things.append(v)
81
+ self.class_names = self.meta.stuff_classes
82
+ self.text_tokenizer = Tokenize(SimpleTokenizer(), max_seq_len=max_seq_len)
83
+ self.task_tokenizer = Tokenize(SimpleTokenizer(), max_seq_len=task_seq_len)
84
+ self.semantic_prob = semantic_prob
85
+ self.instance_prob = instance_prob
86
+
87
+ @classmethod
88
+ def from_config(cls, cfg, is_train=True):
89
+ # Build augmentation
90
+ augs = [
91
+ T.ResizeShortestEdge(
92
+ cfg.INPUT.MIN_SIZE_TRAIN,
93
+ cfg.INPUT.MAX_SIZE_TRAIN,
94
+ cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING,
95
+ )
96
+ ]
97
+ if cfg.INPUT.CROP.ENABLED:
98
+ augs.append(
99
+ T.RandomCrop_CategoryAreaConstraint(
100
+ cfg.INPUT.CROP.TYPE,
101
+ cfg.INPUT.CROP.SIZE,
102
+ cfg.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA,
103
+ cfg.MODEL.SEM_SEG_HEAD.IGNORE_VALUE,
104
+ )
105
+ )
106
+ if cfg.INPUT.COLOR_AUG_SSD:
107
+ augs.append(ColorAugSSDTransform(img_format=cfg.INPUT.FORMAT))
108
+ augs.append(T.RandomFlip())
109
+
110
+ # Assume always applies to the training set.
111
+ dataset_names = cfg.DATASETS.TRAIN
112
+ meta = MetadataCatalog.get(dataset_names[0])
113
+ ignore_label = meta.ignore_label
114
+
115
+ ret = {
116
+ "is_train": is_train,
117
+ "meta": meta,
118
+ "name": dataset_names[0],
119
+ "num_queries": cfg.MODEL.ONE_FORMER.NUM_OBJECT_QUERIES - cfg.MODEL.TEXT_ENCODER.N_CTX,
120
+ "task_seq_len": cfg.INPUT.TASK_SEQ_LEN,
121
+ "max_seq_len": cfg.INPUT.MAX_SEQ_LEN,
122
+ "augmentations": augs,
123
+ "image_format": cfg.INPUT.FORMAT,
124
+ "ignore_label": ignore_label,
125
+ "size_divisibility": cfg.INPUT.SIZE_DIVISIBILITY,
126
+ "semantic_prob": cfg.INPUT.TASK_PROB.SEMANTIC,
127
+ "instance_prob": cfg.INPUT.TASK_PROB.INSTANCE,
128
+ }
129
+ return ret
130
+
131
+ def _get_semantic_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
132
+ pan_seg_gt = pan_seg_gt.numpy()
133
+ instances = Instances(image_shape)
134
+
135
+ classes = []
136
+ texts = ["a semantic photo"] * self.num_queries
137
+ masks = []
138
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
139
+
140
+ for segment_info in segments_info:
141
+ class_id = segment_info["category_id"]
142
+ if not segment_info["iscrowd"]:
143
+ mask = pan_seg_gt == segment_info["id"]
144
+ if not np.all(mask == False):
145
+ if class_id not in classes:
146
+ cls_name = self.class_names[class_id]
147
+ classes.append(class_id)
148
+ masks.append(mask)
149
+ num_class_obj[cls_name] += 1
150
+ else:
151
+ idx = classes.index(class_id)
152
+ masks[idx] += mask
153
+ masks[idx] = np.clip(masks[idx], 0, 1).astype(np.bool)
154
+ label[mask] = class_id
155
+
156
+ num = 0
157
+ for i, cls_name in enumerate(self.class_names):
158
+ if num_class_obj[cls_name] > 0:
159
+ for _ in range(num_class_obj[cls_name]):
160
+ if num >= len(texts):
161
+ break
162
+ texts[num] = f"a photo with a {cls_name}"
163
+ num += 1
164
+
165
+ classes = np.array(classes)
166
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
167
+ if len(masks) == 0:
168
+ # Some image does not have annotation (all ignored)
169
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
170
+ instances.gt_bboxes = torch.zeros((0, 4))
171
+ else:
172
+ masks = BitMasks(
173
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
174
+ )
175
+ instances.gt_masks = masks.tensor
176
+ # Placeholder bounding boxes for stuff regions. Note that these are not used during training.
177
+ instances.gt_bboxes = torch.stack([torch.tensor([0., 0., 1., 1.])] * instances.gt_masks.shape[0])
178
+ return instances, texts, label
179
+
180
+ def _get_instance_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
181
+ pan_seg_gt = pan_seg_gt.numpy()
182
+ instances = Instances(image_shape)
183
+
184
+ classes = []
185
+ texts = ["an instance photo"] * self.num_queries
186
+ masks = []
187
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
188
+
189
+ for segment_info in segments_info:
190
+ class_id = segment_info["category_id"]
191
+ if class_id in self.things:
192
+ if not segment_info["iscrowd"]:
193
+ mask = pan_seg_gt == segment_info["id"]
194
+ if not np.all(mask == False):
195
+ cls_name = self.class_names[class_id]
196
+ classes.append(class_id)
197
+ masks.append(mask)
198
+ num_class_obj[cls_name] += 1
199
+ label[mask] = class_id
200
+
201
+ num = 0
202
+ for i, cls_name in enumerate(self.class_names):
203
+ if num_class_obj[cls_name] > 0:
204
+ for _ in range(num_class_obj[cls_name]):
205
+ if num >= len(texts):
206
+ break
207
+ texts[num] = f"a photo with a {cls_name}"
208
+ num += 1
209
+
210
+ classes = np.array(classes)
211
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
212
+ if len(masks) == 0:
213
+ # Some image does not have annotation (all ignored)
214
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
215
+ instances.gt_bboxes = torch.zeros((0, 4))
216
+ else:
217
+ masks = BitMasks(
218
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
219
+ )
220
+ instances.gt_masks = masks.tensor
221
+ instances.gt_bboxes = masks_to_boxes(instances.gt_masks)
222
+ return instances, texts, label
223
+
224
+ def _get_panoptic_dict(self, pan_seg_gt, image_shape, segments_info, num_class_obj):
225
+ pan_seg_gt = pan_seg_gt.numpy()
226
+ instances = Instances(image_shape)
227
+
228
+ classes = []
229
+ texts = ["a panoptic photo"] * self.num_queries
230
+ masks = []
231
+ label = np.ones_like(pan_seg_gt) * self.ignore_label
232
+
233
+ for segment_info in segments_info:
234
+ class_id = segment_info["category_id"]
235
+ if not segment_info["iscrowd"]:
236
+ mask = pan_seg_gt == segment_info["id"]
237
+ if not np.all(mask == False):
238
+ cls_name = self.class_names[class_id]
239
+ classes.append(class_id)
240
+ masks.append(mask)
241
+ num_class_obj[cls_name] += 1
242
+ label[mask] = class_id
243
+
244
+ num = 0
245
+ for i, cls_name in enumerate(self.class_names):
246
+ if num_class_obj[cls_name] > 0:
247
+ for _ in range(num_class_obj[cls_name]):
248
+ if num >= len(texts):
249
+ break
250
+ texts[num] = f"a photo with a {cls_name}"
251
+ num += 1
252
+
253
+ classes = np.array(classes)
254
+ instances.gt_classes = torch.tensor(classes, dtype=torch.int64)
255
+ if len(masks) == 0:
256
+ # Some image does not have annotation (all ignored)
257
+ instances.gt_masks = torch.zeros((0, pan_seg_gt.shape[-2], pan_seg_gt.shape[-1]))
258
+ instances.gt_bboxes = torch.zeros((0, 4))
259
+ else:
260
+ masks = BitMasks(
261
+ torch.stack([torch.from_numpy(np.ascontiguousarray(x.copy())) for x in masks])
262
+ )
263
+ instances.gt_masks = masks.tensor
264
+ instances.gt_bboxes = masks_to_boxes(instances.gt_masks)
265
+ for i in range(instances.gt_classes.shape[0]):
266
+ # Placeholder bounding boxes for stuff regions. Note that these are not used during training.
267
+ if instances.gt_classes[i].item() not in self.things:
268
+ instances.gt_bboxes[i] = torch.tensor([0., 0., 1., 1.])
269
+ return instances, texts, label
270
+
271
+ def __call__(self, dataset_dict):
272
+ """
273
+ Args:
274
+ dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
275
+
276
+ Returns:
277
+ dict: a format that builtin models in detectron2 accept
278
+ """
279
+ assert self.is_train, "OneFormerUnifiedDatasetMapper should only be used for training!"
280
+
281
+ dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below
282
+ image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
283
+ utils.check_image_size(dataset_dict, image)
284
+
285
+ # semantic segmentation
286
+ if "sem_seg_file_name" in dataset_dict:
287
+ # PyTorch transformation not implemented for uint16, so converting it to double first
288
+ sem_seg_gt = utils.read_image(dataset_dict.pop("sem_seg_file_name")).astype("double")
289
+ else:
290
+ sem_seg_gt = None
291
+
292
+ # panoptic segmentation
293
+ if "pan_seg_file_name" in dataset_dict:
294
+ pan_seg_gt = utils.read_image(dataset_dict.pop("pan_seg_file_name"), "RGB")
295
+ segments_info = dataset_dict["segments_info"]
296
+ else:
297
+ pan_seg_gt = None
298
+ segments_info = None
299
+
300
+ if pan_seg_gt is None:
301
+ raise ValueError(
302
+ "Cannot find 'pan_seg_file_name' for panoptic segmentation dataset {}.".format(
303
+ dataset_dict["file_name"]
304
+ )
305
+ )
306
+
307
+ aug_input = T.AugInput(image, sem_seg=sem_seg_gt)
308
+ aug_input, transforms = T.apply_transform_gens(self.tfm_gens, aug_input)
309
+ image = aug_input.image
310
+ if sem_seg_gt is not None:
311
+ sem_seg_gt = aug_input.sem_seg
312
+
313
+ # apply the same transformation to panoptic segmentation
314
+ pan_seg_gt = transforms.apply_segmentation(pan_seg_gt)
315
+
316
+ from panopticapi.utils import rgb2id
317
+
318
+ pan_seg_gt = rgb2id(pan_seg_gt)
319
+
320
+ # Pad image and segmentation label here!
321
+ image = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
322
+ if sem_seg_gt is not None:
323
+ sem_seg_gt = torch.as_tensor(sem_seg_gt.astype("long"))
324
+ pan_seg_gt = torch.as_tensor(pan_seg_gt.astype("long"))
325
+
326
+ if self.size_divisibility > 0:
327
+ image_size = (image.shape[-2], image.shape[-1])
328
+ padding_size = [
329
+ 0,
330
+ self.size_divisibility - image_size[1],
331
+ 0,
332
+ self.size_divisibility - image_size[0],
333
+ ]
334
+ image = F.pad(image, padding_size, value=128).contiguous()
335
+ if sem_seg_gt is not None:
336
+ sem_seg_gt = F.pad(sem_seg_gt, padding_size, value=self.ignore_label).contiguous()
337
+ pan_seg_gt = F.pad(
338
+ pan_seg_gt, padding_size, value=0
339
+ ).contiguous() # 0 is the VOID panoptic label
340
+
341
+ image_shape = (image.shape[-2], image.shape[-1]) # h, w
342
+
343
+ # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
344
+ # but not efficient on large generic data structures due to the use of pickle & mp.Queue.
345
+ # Therefore it's important to use torch.Tensor.
346
+ dataset_dict["image"] = image
347
+
348
+ if "annotations" in dataset_dict:
349
+ raise ValueError("Pemantic segmentation dataset should not have 'annotations'.")
350
+
351
+ prob_task = np.random.uniform(0,1.)
352
+
353
+ num_class_obj = {}
354
+
355
+ for name in self.class_names:
356
+ num_class_obj[name] = 0
357
+
358
+ if prob_task < self.semantic_prob:
359
+ task = "The task is semantic"
360
+ instances, text, sem_seg = self._get_semantic_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
361
+ elif prob_task < self.instance_prob:
362
+ task = "The task is instance"
363
+ instances, text, sem_seg = self._get_instance_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
364
+ else:
365
+ task = "The task is panoptic"
366
+ instances, text, sem_seg = self._get_panoptic_dict(pan_seg_gt, image_shape, segments_info, num_class_obj)
367
+
368
+ dataset_dict["sem_seg"] = torch.from_numpy(sem_seg).long()
369
+ dataset_dict["instances"] = instances
370
+ dataset_dict["orig_shape"] = image_shape
371
+ dataset_dict["task"] = task
372
+ dataset_dict["text"] = text
373
+ dataset_dict["thing_ids"] = self.things
374
+
375
+ return dataset_dict
oneformer/data/datasets/__init__.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ from . import (
2
+ register_ade20k_panoptic,
3
+ register_cityscapes_panoptic,
4
+ register_coco_panoptic_annos_semseg,
5
+ register_ade20k_instance,
6
+ register_coco_panoptic2instance,
7
+ )
oneformer/data/datasets/register_ade20k_instance.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/data/datasets/register_ade20k_instance.py
3
+ # ------------------------------------------------------------------------------
4
+
5
+ import json
6
+ import logging
7
+ import numpy as np
8
+ import os
9
+ from PIL import Image
10
+
11
+ from detectron2.data import DatasetCatalog, MetadataCatalog
12
+ from detectron2.data.datasets.coco import load_coco_json, register_coco_instances
13
+ from detectron2.utils.file_io import PathManager
14
+
15
+ ADE_CATEGORIES = [{'id': 7, 'name': 'bed'}, {'id': 8, 'name': 'windowpane'}, {'id': 10, 'name': 'cabinet'}, {'id': 12, 'name': 'person'}, {'id': 14, 'name': 'door'}, {'id': 15, 'name': 'table'}, {'id': 18, 'name': 'curtain'}, {'id': 19, 'name': 'chair'}, {'id': 20, 'name': 'car'}, {'id': 22, 'name': 'painting'}, {'id': 23, 'name': 'sofa'}, {'id': 24, 'name': 'shelf'}, {'id': 27, 'name': 'mirror'}, {'id': 30, 'name': 'armchair'}, {'id': 31, 'name': 'seat'}, {'id': 32, 'name': 'fence'}, {'id': 33, 'name': 'desk'}, {'id': 35, 'name': 'wardrobe'}, {'id': 36, 'name': 'lamp'}, {'id': 37, 'name': 'bathtub'}, {'id': 38, 'name': 'railing'}, {'id': 39, 'name': 'cushion'}, {'id': 41, 'name': 'box'}, {'id': 42, 'name': 'column'}, {'id': 43, 'name': 'signboard'}, {'id': 44, 'name': 'chest of drawers'}, {'id': 45, 'name': 'counter'}, {'id': 47, 'name': 'sink'}, {'id': 49, 'name': 'fireplace'}, {'id': 50, 'name': 'refrigerator'}, {'id': 53, 'name': 'stairs'}, {'id': 55, 'name': 'case'}, {'id': 56, 'name': 'pool table'}, {'id': 57, 'name': 'pillow'}, {'id': 58, 'name': 'screen door'}, {'id': 62, 'name': 'bookcase'}, {'id': 64, 'name': 'coffee table'}, {'id': 65, 'name': 'toilet'}, {'id': 66, 'name': 'flower'}, {'id': 67, 'name': 'book'}, {'id': 69, 'name': 'bench'}, {'id': 70, 'name': 'countertop'}, {'id': 71, 'name': 'stove'}, {'id': 72, 'name': 'palm'}, {'id': 73, 'name': 'kitchen island'}, {'id': 74, 'name': 'computer'}, {'id': 75, 'name': 'swivel chair'}, {'id': 76, 'name': 'boat'}, {'id': 78, 'name': 'arcade machine'}, {'id': 80, 'name': 'bus'}, {'id': 81, 'name': 'towel'}, {'id': 82, 'name': 'light'}, {'id': 83, 'name': 'truck'}, {'id': 85, 'name': 'chandelier'}, {'id': 86, 'name': 'awning'}, {'id': 87, 'name': 'streetlight'}, {'id': 88, 'name': 'booth'}, {'id': 89, 'name': 'television receiver'}, {'id': 90, 'name': 'airplane'}, {'id': 92, 'name': 'apparel'}, {'id': 93, 'name': 'pole'}, {'id': 95, 'name': 'bannister'}, {'id': 97, 'name': 'ottoman'}, {'id': 98, 'name': 'bottle'}, {'id': 102, 'name': 'van'}, {'id': 103, 'name': 'ship'}, {'id': 104, 'name': 'fountain'}, {'id': 107, 'name': 'washer'}, {'id': 108, 'name': 'plaything'}, {'id': 110, 'name': 'stool'}, {'id': 111, 'name': 'barrel'}, {'id': 112, 'name': 'basket'}, {'id': 115, 'name': 'bag'}, {'id': 116, 'name': 'minibike'}, {'id': 118, 'name': 'oven'}, {'id': 119, 'name': 'ball'}, {'id': 120, 'name': 'food'}, {'id': 121, 'name': 'step'}, {'id': 123, 'name': 'trade name'}, {'id': 124, 'name': 'microwave'}, {'id': 125, 'name': 'pot'}, {'id': 126, 'name': 'animal'}, {'id': 127, 'name': 'bicycle'}, {'id': 129, 'name': 'dishwasher'}, {'id': 130, 'name': 'screen'}, {'id': 132, 'name': 'sculpture'}, {'id': 133, 'name': 'hood'}, {'id': 134, 'name': 'sconce'}, {'id': 135, 'name': 'vase'}, {'id': 136, 'name': 'traffic light'}, {'id': 137, 'name': 'tray'}, {'id': 138, 'name': 'ashcan'}, {'id': 139, 'name': 'fan'}, {'id': 142, 'name': 'plate'}, {'id': 143, 'name': 'monitor'}, {'id': 144, 'name': 'bulletin board'}, {'id': 146, 'name': 'radiator'}, {'id': 147, 'name': 'glass'}, {'id': 148, 'name': 'clock'}, {'id': 149, 'name': 'flag'}]
16
+
17
+
18
+ _PREDEFINED_SPLITS = {
19
+ # point annotations without masks
20
+ "ade20k_instance_train": (
21
+ "ADEChallengeData2016/images/training",
22
+ "ADEChallengeData2016/ade20k_instance_train.json",
23
+ ),
24
+ "ade20k_instance_val": (
25
+ "ADEChallengeData2016/images/validation",
26
+ "ADEChallengeData2016/ade20k_instance_val.json",
27
+ ),
28
+ }
29
+
30
+
31
+ def _get_ade_instances_meta():
32
+ thing_ids = [k["id"] for k in ADE_CATEGORIES]
33
+ assert len(thing_ids) == 100, len(thing_ids)
34
+ # Mapping from the incontiguous ADE category id to an id in [0, 99]
35
+ thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}
36
+ thing_classes = [k["name"] for k in ADE_CATEGORIES]
37
+ ret = {
38
+ "thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id,
39
+ "thing_classes": thing_classes,
40
+ }
41
+ return ret
42
+
43
+
44
+ def register_all_ade20k_instance(root):
45
+ for key, (image_root, json_file) in _PREDEFINED_SPLITS.items():
46
+ # Assume pre-defined datasets live in `./datasets`.
47
+ register_coco_instances(
48
+ key,
49
+ _get_ade_instances_meta(),
50
+ os.path.join(root, json_file) if "://" not in json_file else json_file,
51
+ os.path.join(root, image_root),
52
+ )
53
+
54
+
55
+ _root = os.getenv("DETECTRON2_DATASETS", "datasets")
56
+ register_all_ade20k_instance(_root)
oneformer/data/datasets/register_ade20k_panoptic.py ADDED
@@ -0,0 +1,394 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/data/datasets/register_ade20k_panoptic.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import json
7
+ import os
8
+
9
+ from detectron2.data import DatasetCatalog, MetadataCatalog
10
+ from detectron2.utils.file_io import PathManager
11
+
12
+ ADE20K_150_CATEGORIES = [
13
+ {"color": [120, 120, 120], "id": 0, "isthing": 0, "name": "wall"},
14
+ {"color": [180, 120, 120], "id": 1, "isthing": 0, "name": "building"},
15
+ {"color": [6, 230, 230], "id": 2, "isthing": 0, "name": "sky"},
16
+ {"color": [80, 50, 50], "id": 3, "isthing": 0, "name": "floor"},
17
+ {"color": [4, 200, 3], "id": 4, "isthing": 0, "name": "tree"},
18
+ {"color": [120, 120, 80], "id": 5, "isthing": 0, "name": "ceiling"},
19
+ {"color": [140, 140, 140], "id": 6, "isthing": 0, "name": "road, route"},
20
+ {"color": [204, 5, 255], "id": 7, "isthing": 1, "name": "bed"},
21
+ {"color": [230, 230, 230], "id": 8, "isthing": 1, "name": "window "},
22
+ {"color": [4, 250, 7], "id": 9, "isthing": 0, "name": "grass"},
23
+ {"color": [224, 5, 255], "id": 10, "isthing": 1, "name": "cabinet"},
24
+ {"color": [235, 255, 7], "id": 11, "isthing": 0, "name": "sidewalk, pavement"},
25
+ {"color": [150, 5, 61], "id": 12, "isthing": 1, "name": "person"},
26
+ {"color": [120, 120, 70], "id": 13, "isthing": 0, "name": "earth, ground"},
27
+ {"color": [8, 255, 51], "id": 14, "isthing": 1, "name": "door"},
28
+ {"color": [255, 6, 82], "id": 15, "isthing": 1, "name": "table"},
29
+ {"color": [143, 255, 140], "id": 16, "isthing": 0, "name": "mountain, mount"},
30
+ {"color": [204, 255, 4], "id": 17, "isthing": 0, "name": "plant"},
31
+ {"color": [255, 51, 7], "id": 18, "isthing": 1, "name": "curtain"},
32
+ {"color": [204, 70, 3], "id": 19, "isthing": 1, "name": "chair"},
33
+ {"color": [0, 102, 200], "id": 20, "isthing": 1, "name": "car"},
34
+ {"color": [61, 230, 250], "id": 21, "isthing": 0, "name": "water"},
35
+ {"color": [255, 6, 51], "id": 22, "isthing": 1, "name": "painting, picture"},
36
+ {"color": [11, 102, 255], "id": 23, "isthing": 1, "name": "sofa"},
37
+ {"color": [255, 7, 71], "id": 24, "isthing": 1, "name": "shelf"},
38
+ {"color": [255, 9, 224], "id": 25, "isthing": 0, "name": "house"},
39
+ {"color": [9, 7, 230], "id": 26, "isthing": 0, "name": "sea"},
40
+ {"color": [220, 220, 220], "id": 27, "isthing": 1, "name": "mirror"},
41
+ {"color": [255, 9, 92], "id": 28, "isthing": 0, "name": "rug"},
42
+ {"color": [112, 9, 255], "id": 29, "isthing": 0, "name": "field"},
43
+ {"color": [8, 255, 214], "id": 30, "isthing": 1, "name": "armchair"},
44
+ {"color": [7, 255, 224], "id": 31, "isthing": 1, "name": "seat"},
45
+ {"color": [255, 184, 6], "id": 32, "isthing": 1, "name": "fence"},
46
+ {"color": [10, 255, 71], "id": 33, "isthing": 1, "name": "desk"},
47
+ {"color": [255, 41, 10], "id": 34, "isthing": 0, "name": "rock, stone"},
48
+ {"color": [7, 255, 255], "id": 35, "isthing": 1, "name": "wardrobe, closet, press"},
49
+ {"color": [224, 255, 8], "id": 36, "isthing": 1, "name": "lamp"},
50
+ {"color": [102, 8, 255], "id": 37, "isthing": 1, "name": "tub"},
51
+ {"color": [255, 61, 6], "id": 38, "isthing": 1, "name": "rail"},
52
+ {"color": [255, 194, 7], "id": 39, "isthing": 1, "name": "cushion"},
53
+ {"color": [255, 122, 8], "id": 40, "isthing": 0, "name": "base, pedestal, stand"},
54
+ {"color": [0, 255, 20], "id": 41, "isthing": 1, "name": "box"},
55
+ {"color": [255, 8, 41], "id": 42, "isthing": 1, "name": "column, pillar"},
56
+ {"color": [255, 5, 153], "id": 43, "isthing": 1, "name": "signboard, sign"},
57
+ {
58
+ "color": [6, 51, 255],
59
+ "id": 44,
60
+ "isthing": 1,
61
+ "name": "chest of drawers, chest, bureau, dresser",
62
+ },
63
+ {"color": [235, 12, 255], "id": 45, "isthing": 1, "name": "counter"},
64
+ {"color": [160, 150, 20], "id": 46, "isthing": 0, "name": "sand"},
65
+ {"color": [0, 163, 255], "id": 47, "isthing": 1, "name": "sink"},
66
+ {"color": [140, 140, 140], "id": 48, "isthing": 0, "name": "skyscraper"},
67
+ {"color": [250, 10, 15], "id": 49, "isthing": 1, "name": "fireplace"},
68
+ {"color": [20, 255, 0], "id": 50, "isthing": 1, "name": "refrigerator, icebox"},
69
+ {"color": [31, 255, 0], "id": 51, "isthing": 0, "name": "grandstand, covered stand"},
70
+ {"color": [255, 31, 0], "id": 52, "isthing": 0, "name": "path"},
71
+ {"color": [255, 224, 0], "id": 53, "isthing": 1, "name": "stairs"},
72
+ {"color": [153, 255, 0], "id": 54, "isthing": 0, "name": "runway"},
73
+ {"color": [0, 0, 255], "id": 55, "isthing": 1, "name": "case, display case, showcase, vitrine"},
74
+ {
75
+ "color": [255, 71, 0],
76
+ "id": 56,
77
+ "isthing": 1,
78
+ "name": "pool table, billiard table, snooker table",
79
+ },
80
+ {"color": [0, 235, 255], "id": 57, "isthing": 1, "name": "pillow"},
81
+ {"color": [0, 173, 255], "id": 58, "isthing": 1, "name": "screen door, screen"},
82
+ {"color": [31, 0, 255], "id": 59, "isthing": 0, "name": "stairway, staircase"},
83
+ {"color": [11, 200, 200], "id": 60, "isthing": 0, "name": "river"},
84
+ {"color": [255, 82, 0], "id": 61, "isthing": 0, "name": "bridge, span"},
85
+ {"color": [0, 255, 245], "id": 62, "isthing": 1, "name": "bookcase"},
86
+ {"color": [0, 61, 255], "id": 63, "isthing": 0, "name": "blind, screen"},
87
+ {"color": [0, 255, 112], "id": 64, "isthing": 1, "name": "coffee table"},
88
+ {
89
+ "color": [0, 255, 133],
90
+ "id": 65,
91
+ "isthing": 1,
92
+ "name": "toilet, can, commode, crapper, pot, potty, stool, throne",
93
+ },
94
+ {"color": [255, 0, 0], "id": 66, "isthing": 1, "name": "flower"},
95
+ {"color": [255, 163, 0], "id": 67, "isthing": 1, "name": "book"},
96
+ {"color": [255, 102, 0], "id": 68, "isthing": 0, "name": "hill"},
97
+ {"color": [194, 255, 0], "id": 69, "isthing": 1, "name": "bench"},
98
+ {"color": [0, 143, 255], "id": 70, "isthing": 1, "name": "countertop"},
99
+ {"color": [51, 255, 0], "id": 71, "isthing": 1, "name": "stove"},
100
+ {"color": [0, 82, 255], "id": 72, "isthing": 1, "name": "palm, palm tree"},
101
+ {"color": [0, 255, 41], "id": 73, "isthing": 1, "name": "kitchen island"},
102
+ {"color": [0, 255, 173], "id": 74, "isthing": 1, "name": "computer"},
103
+ {"color": [10, 0, 255], "id": 75, "isthing": 1, "name": "swivel chair"},
104
+ {"color": [173, 255, 0], "id": 76, "isthing": 1, "name": "boat"},
105
+ {"color": [0, 255, 153], "id": 77, "isthing": 0, "name": "bar"},
106
+ {"color": [255, 92, 0], "id": 78, "isthing": 1, "name": "arcade machine"},
107
+ {"color": [255, 0, 255], "id": 79, "isthing": 0, "name": "hovel, hut, hutch, shack, shanty"},
108
+ {"color": [255, 0, 245], "id": 80, "isthing": 1, "name": "bus"},
109
+ {"color": [255, 0, 102], "id": 81, "isthing": 1, "name": "towel"},
110
+ {"color": [255, 173, 0], "id": 82, "isthing": 1, "name": "light"},
111
+ {"color": [255, 0, 20], "id": 83, "isthing": 1, "name": "truck"},
112
+ {"color": [255, 184, 184], "id": 84, "isthing": 0, "name": "tower"},
113
+ {"color": [0, 31, 255], "id": 85, "isthing": 1, "name": "chandelier"},
114
+ {"color": [0, 255, 61], "id": 86, "isthing": 1, "name": "awning, sunshade, sunblind"},
115
+ {"color": [0, 71, 255], "id": 87, "isthing": 1, "name": "street lamp"},
116
+ {"color": [255, 0, 204], "id": 88, "isthing": 1, "name": "booth"},
117
+ {"color": [0, 255, 194], "id": 89, "isthing": 1, "name": "tv"},
118
+ {"color": [0, 255, 82], "id": 90, "isthing": 1, "name": "plane"},
119
+ {"color": [0, 10, 255], "id": 91, "isthing": 0, "name": "dirt track"},
120
+ {"color": [0, 112, 255], "id": 92, "isthing": 1, "name": "clothes"},
121
+ {"color": [51, 0, 255], "id": 93, "isthing": 1, "name": "pole"},
122
+ {"color": [0, 194, 255], "id": 94, "isthing": 0, "name": "land, ground, soil"},
123
+ {
124
+ "color": [0, 122, 255],
125
+ "id": 95,
126
+ "isthing": 1,
127
+ "name": "bannister, banister, balustrade, balusters, handrail",
128
+ },
129
+ {
130
+ "color": [0, 255, 163],
131
+ "id": 96,
132
+ "isthing": 0,
133
+ "name": "escalator, moving staircase, moving stairway",
134
+ },
135
+ {
136
+ "color": [255, 153, 0],
137
+ "id": 97,
138
+ "isthing": 1,
139
+ "name": "ottoman, pouf, pouffe, puff, hassock",
140
+ },
141
+ {"color": [0, 255, 10], "id": 98, "isthing": 1, "name": "bottle"},
142
+ {"color": [255, 112, 0], "id": 99, "isthing": 0, "name": "buffet, counter, sideboard"},
143
+ {
144
+ "color": [143, 255, 0],
145
+ "id": 100,
146
+ "isthing": 0,
147
+ "name": "poster, posting, placard, notice, bill, card",
148
+ },
149
+ {"color": [82, 0, 255], "id": 101, "isthing": 0, "name": "stage"},
150
+ {"color": [163, 255, 0], "id": 102, "isthing": 1, "name": "van"},
151
+ {"color": [255, 235, 0], "id": 103, "isthing": 1, "name": "ship"},
152
+ {"color": [8, 184, 170], "id": 104, "isthing": 1, "name": "fountain"},
153
+ {
154
+ "color": [133, 0, 255],
155
+ "id": 105,
156
+ "isthing": 0,
157
+ "name": "conveyer belt, conveyor belt, conveyer, conveyor, transporter",
158
+ },
159
+ {"color": [0, 255, 92], "id": 106, "isthing": 0, "name": "canopy"},
160
+ {
161
+ "color": [184, 0, 255],
162
+ "id": 107,
163
+ "isthing": 1,
164
+ "name": "washer, automatic washer, washing machine",
165
+ },
166
+ {"color": [255, 0, 31], "id": 108, "isthing": 1, "name": "plaything, toy"},
167
+ {"color": [0, 184, 255], "id": 109, "isthing": 0, "name": "pool"},
168
+ {"color": [0, 214, 255], "id": 110, "isthing": 1, "name": "stool"},
169
+ {"color": [255, 0, 112], "id": 111, "isthing": 1, "name": "barrel, cask"},
170
+ {"color": [92, 255, 0], "id": 112, "isthing": 1, "name": "basket, handbasket"},
171
+ {"color": [0, 224, 255], "id": 113, "isthing": 0, "name": "falls"},
172
+ {"color": [112, 224, 255], "id": 114, "isthing": 0, "name": "tent"},
173
+ {"color": [70, 184, 160], "id": 115, "isthing": 1, "name": "bag"},
174
+ {"color": [163, 0, 255], "id": 116, "isthing": 1, "name": "minibike, motorbike"},
175
+ {"color": [153, 0, 255], "id": 117, "isthing": 0, "name": "cradle"},
176
+ {"color": [71, 255, 0], "id": 118, "isthing": 1, "name": "oven"},
177
+ {"color": [255, 0, 163], "id": 119, "isthing": 1, "name": "ball"},
178
+ {"color": [255, 204, 0], "id": 120, "isthing": 1, "name": "food, solid food"},
179
+ {"color": [255, 0, 143], "id": 121, "isthing": 1, "name": "step, stair"},
180
+ {"color": [0, 255, 235], "id": 122, "isthing": 0, "name": "tank, storage tank"},
181
+ {"color": [133, 255, 0], "id": 123, "isthing": 1, "name": "trade name"},
182
+ {"color": [255, 0, 235], "id": 124, "isthing": 1, "name": "microwave"},
183
+ {"color": [245, 0, 255], "id": 125, "isthing": 1, "name": "pot"},
184
+ {"color": [255, 0, 122], "id": 126, "isthing": 1, "name": "animal"},
185
+ {"color": [255, 245, 0], "id": 127, "isthing": 1, "name": "bicycle"},
186
+ {"color": [10, 190, 212], "id": 128, "isthing": 0, "name": "lake"},
187
+ {"color": [214, 255, 0], "id": 129, "isthing": 1, "name": "dishwasher"},
188
+ {"color": [0, 204, 255], "id": 130, "isthing": 1, "name": "screen"},
189
+ {"color": [20, 0, 255], "id": 131, "isthing": 0, "name": "blanket, cover"},
190
+ {"color": [255, 255, 0], "id": 132, "isthing": 1, "name": "sculpture"},
191
+ {"color": [0, 153, 255], "id": 133, "isthing": 1, "name": "hood, exhaust hood"},
192
+ {"color": [0, 41, 255], "id": 134, "isthing": 1, "name": "sconce"},
193
+ {"color": [0, 255, 204], "id": 135, "isthing": 1, "name": "vase"},
194
+ {"color": [41, 0, 255], "id": 136, "isthing": 1, "name": "traffic light"},
195
+ {"color": [41, 255, 0], "id": 137, "isthing": 1, "name": "tray"},
196
+ {"color": [173, 0, 255], "id": 138, "isthing": 1, "name": "trash can"},
197
+ {"color": [0, 245, 255], "id": 139, "isthing": 1, "name": "fan"},
198
+ {"color": [71, 0, 255], "id": 140, "isthing": 0, "name": "pier"},
199
+ {"color": [122, 0, 255], "id": 141, "isthing": 0, "name": "crt screen"},
200
+ {"color": [0, 255, 184], "id": 142, "isthing": 1, "name": "plate"},
201
+ {"color": [0, 92, 255], "id": 143, "isthing": 1, "name": "monitor"},
202
+ {"color": [184, 255, 0], "id": 144, "isthing": 1, "name": "bulletin board"},
203
+ {"color": [0, 133, 255], "id": 145, "isthing": 0, "name": "shower"},
204
+ {"color": [255, 214, 0], "id": 146, "isthing": 1, "name": "radiator"},
205
+ {"color": [25, 194, 194], "id": 147, "isthing": 1, "name": "glass, drinking glass"},
206
+ {"color": [102, 255, 0], "id": 148, "isthing": 1, "name": "clock"},
207
+ {"color": [92, 0, 255], "id": 149, "isthing": 1, "name": "flag"},
208
+ ]
209
+
210
+ ADE20k_COLORS = [k["color"] for k in ADE20K_150_CATEGORIES]
211
+
212
+ MetadataCatalog.get("ade20k_sem_seg_train").set(
213
+ stuff_colors=ADE20k_COLORS[:],
214
+ )
215
+
216
+ MetadataCatalog.get("ade20k_sem_seg_val").set(
217
+ stuff_colors=ADE20k_COLORS[:],
218
+ )
219
+
220
+
221
+ def load_ade20k_panoptic_json(json_file, image_dir, gt_dir, semseg_dir, meta):
222
+ """
223
+ Args:
224
+ image_dir (str): path to the raw dataset. e.g., "~/coco/train2017".
225
+ gt_dir (str): path to the raw annotations. e.g., "~/coco/panoptic_train2017".
226
+ json_file (str): path to the json file. e.g., "~/coco/annotations/panoptic_train2017.json".
227
+ Returns:
228
+ list[dict]: a list of dicts in Detectron2 standard format. (See
229
+ `Using Custom Datasets </tutorials/datasets.html>`_ )
230
+ """
231
+
232
+ def _convert_category_id(segment_info, meta):
233
+ if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]:
234
+ segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][
235
+ segment_info["category_id"]
236
+ ]
237
+ segment_info["isthing"] = True
238
+ else:
239
+ segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][
240
+ segment_info["category_id"]
241
+ ]
242
+ segment_info["isthing"] = False
243
+ return segment_info
244
+
245
+ with PathManager.open(json_file) as f:
246
+ json_info = json.load(f)
247
+
248
+ ret = []
249
+ for ann in json_info["annotations"]:
250
+ image_id = ann["image_id"]
251
+ # TODO: currently we assume image and label has the same filename but
252
+ # different extension, and images have extension ".jpg" for COCO. Need
253
+ # to make image extension a user-provided argument if we extend this
254
+ # function to support other COCO-like datasets.
255
+ image_file = os.path.join(image_dir, os.path.splitext(ann["file_name"])[0] + ".jpg")
256
+ label_file = os.path.join(gt_dir, ann["file_name"])
257
+ sem_label_file = os.path.join(semseg_dir, ann["file_name"])
258
+ segments_info = [_convert_category_id(x, meta) for x in ann["segments_info"]]
259
+ ret.append(
260
+ {
261
+ "file_name": image_file,
262
+ "image_id": image_id,
263
+ "pan_seg_file_name": label_file,
264
+ "sem_seg_file_name": sem_label_file,
265
+ "segments_info": segments_info,
266
+ }
267
+ )
268
+ assert len(ret), f"No images found in {image_dir}!"
269
+ assert PathManager.isfile(ret[0]["file_name"]), ret[0]["file_name"]
270
+ assert PathManager.isfile(ret[0]["pan_seg_file_name"]), ret[0]["pan_seg_file_name"]
271
+ assert PathManager.isfile(ret[0]["sem_seg_file_name"]), ret[0]["sem_seg_file_name"]
272
+ return ret
273
+
274
+
275
+ def register_ade20k_panoptic(
276
+ name, metadata, image_root, panoptic_root, semantic_root, panoptic_json, instances_json=None,
277
+ ):
278
+ """
279
+ Register a "standard" version of ADE20k panoptic segmentation dataset named `name`.
280
+ The dictionaries in this registered dataset follows detectron2's standard format.
281
+ Hence it's called "standard".
282
+ Args:
283
+ name (str): the name that identifies a dataset,
284
+ e.g. "ade20k_panoptic_train"
285
+ metadata (dict): extra metadata associated with this dataset.
286
+ image_root (str): directory which contains all the images
287
+ panoptic_root (str): directory which contains panoptic annotation images in COCO format
288
+ panoptic_json (str): path to the json panoptic annotation file in COCO format
289
+ sem_seg_root (none): not used, to be consistent with
290
+ `register_coco_panoptic_separated`.
291
+ instances_json (str): path to the json instance annotation file
292
+ """
293
+ panoptic_name = name
294
+ DatasetCatalog.register(
295
+ panoptic_name,
296
+ lambda: load_ade20k_panoptic_json(
297
+ panoptic_json, image_root, panoptic_root, semantic_root, metadata
298
+ ),
299
+ )
300
+ MetadataCatalog.get(panoptic_name).set(
301
+ panoptic_root=panoptic_root,
302
+ image_root=image_root,
303
+ panoptic_json=panoptic_json,
304
+ json_file=instances_json,
305
+ evaluator_type="ade20k_panoptic_seg",
306
+ ignore_label=255,
307
+ label_divisor=1000,
308
+ **metadata,
309
+ )
310
+
311
+
312
+ _PREDEFINED_SPLITS_ADE20K_PANOPTIC = {
313
+ "ade20k_panoptic_train": (
314
+ "ADEChallengeData2016/images/training",
315
+ "ADEChallengeData2016/ade20k_panoptic_train",
316
+ "ADEChallengeData2016/ade20k_panoptic_train.json",
317
+ "ADEChallengeData2016/annotations_detectron2/training",
318
+ "ADEChallengeData2016/ade20k_instance_train.json",
319
+ ),
320
+ "ade20k_panoptic_val": (
321
+ "ADEChallengeData2016/images/validation",
322
+ "ADEChallengeData2016/ade20k_panoptic_val",
323
+ "ADEChallengeData2016/ade20k_panoptic_val.json",
324
+ "ADEChallengeData2016/annotations_detectron2/validation",
325
+ "ADEChallengeData2016/ade20k_instance_val.json",
326
+ ),
327
+ }
328
+
329
+
330
+ def get_metadata():
331
+ meta = {}
332
+ # The following metadata maps contiguous id from [0, #thing categories +
333
+ # #stuff categories) to their names and colors. We have to replica of the
334
+ # same name and color under "thing_*" and "stuff_*" because the current
335
+ # visualization function in D2 handles thing and class classes differently
336
+ # due to some heuristic used in Panoptic FPN. We keep the same naming to
337
+ # enable reusing existing visualization functions.
338
+ thing_classes = [k["name"] for k in ADE20K_150_CATEGORIES if k["isthing"] == 1]
339
+ thing_colors = [k["color"] for k in ADE20K_150_CATEGORIES if k["isthing"] == 1]
340
+ stuff_classes = [k["name"] for k in ADE20K_150_CATEGORIES]
341
+ stuff_colors = [k["color"] for k in ADE20K_150_CATEGORIES]
342
+
343
+ meta["thing_classes"] = thing_classes
344
+ meta["thing_colors"] = thing_colors
345
+ meta["stuff_classes"] = stuff_classes
346
+ meta["stuff_colors"] = stuff_colors
347
+
348
+ # Convert category id for training:
349
+ # category id: like semantic segmentation, it is the class id for each
350
+ # pixel. Since there are some classes not used in evaluation, the category
351
+ # id is not always contiguous and thus we have two set of category ids:
352
+ # - original category id: category id in the original dataset, mainly
353
+ # used for evaluation.
354
+ # - contiguous category id: [0, #classes), in order to train the linear
355
+ # softmax classifier.
356
+ thing_dataset_id_to_contiguous_id = {}
357
+ stuff_dataset_id_to_contiguous_id = {}
358
+
359
+ for i, cat in enumerate(ADE20K_150_CATEGORIES):
360
+ if cat["isthing"]:
361
+ thing_dataset_id_to_contiguous_id[cat["id"]] = i
362
+ # else:
363
+ # stuff_dataset_id_to_contiguous_id[cat["id"]] = i
364
+
365
+ # in order to use sem_seg evaluator
366
+ stuff_dataset_id_to_contiguous_id[cat["id"]] = i
367
+
368
+ meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id
369
+ meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id
370
+
371
+ return meta
372
+
373
+
374
+ def register_all_ade20k_panoptic(root):
375
+ metadata = get_metadata()
376
+ for (
377
+ prefix,
378
+ (image_root, panoptic_root, panoptic_json, semantic_root, instance_json),
379
+ ) in _PREDEFINED_SPLITS_ADE20K_PANOPTIC.items():
380
+ # The "standard" version of COCO panoptic segmentation dataset,
381
+ # e.g. used by Panoptic-DeepLab
382
+ register_ade20k_panoptic(
383
+ prefix,
384
+ metadata,
385
+ os.path.join(root, image_root),
386
+ os.path.join(root, panoptic_root),
387
+ os.path.join(root, semantic_root),
388
+ os.path.join(root, panoptic_json),
389
+ os.path.join(root, instance_json),
390
+ )
391
+
392
+
393
+ _root = os.getenv("DETECTRON2_DATASETS", "datasets")
394
+ register_all_ade20k_panoptic(_root)
oneformer/data/datasets/register_cityscapes_panoptic.py ADDED
@@ -0,0 +1,199 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/datasets/cityscapes_panoptic.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import json
7
+ import logging
8
+ import os
9
+
10
+ from detectron2.data import DatasetCatalog, MetadataCatalog
11
+ from detectron2.data.datasets.builtin_meta import CITYSCAPES_CATEGORIES
12
+ from detectron2.utils.file_io import PathManager
13
+
14
+ """
15
+ This file contains functions to register the Cityscapes panoptic dataset to the DatasetCatalog.
16
+ """
17
+
18
+
19
+ logger = logging.getLogger(__name__)
20
+
21
+
22
+ def get_cityscapes_panoptic_files(image_dir, gt_dir, json_info):
23
+ files = []
24
+ # scan through the directory
25
+ cities = PathManager.ls(image_dir)
26
+ logger.info(f"{len(cities)} cities found in '{image_dir}'.")
27
+ image_dict = {}
28
+ for city in cities:
29
+ city_img_dir = os.path.join(image_dir, city)
30
+ for basename in PathManager.ls(city_img_dir):
31
+ image_file = os.path.join(city_img_dir, basename)
32
+
33
+ suffix = "_leftImg8bit.png"
34
+ assert basename.endswith(suffix), basename
35
+ basename = os.path.basename(basename)[: -len(suffix)]
36
+
37
+ image_dict[basename] = image_file
38
+
39
+ for ann in json_info["annotations"]:
40
+ image_file = image_dict.get(ann["image_id"], None)
41
+ assert image_file is not None, "No image {} found for annotation {}".format(
42
+ ann["image_id"], ann["file_name"]
43
+ )
44
+ label_file = os.path.join(gt_dir, ann["file_name"])
45
+ segments_info = ann["segments_info"]
46
+ files.append((image_file, label_file, segments_info))
47
+
48
+ assert len(files), "No images found in {}".format(image_dir)
49
+ assert PathManager.isfile(files[0][0]), files[0][0]
50
+ assert PathManager.isfile(files[0][1]), files[0][1]
51
+ return files
52
+
53
+
54
+ def load_cityscapes_panoptic(image_dir, gt_dir, gt_json, meta):
55
+ """
56
+ Args:
57
+ image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train".
58
+ gt_dir (str): path to the raw annotations. e.g.,
59
+ "~/cityscapes/gtFine/cityscapes_panoptic_train".
60
+ gt_json (str): path to the json file. e.g.,
61
+ "~/cityscapes/gtFine/cityscapes_panoptic_train.json".
62
+ meta (dict): dictionary containing "thing_dataset_id_to_contiguous_id"
63
+ and "stuff_dataset_id_to_contiguous_id" to map category ids to
64
+ contiguous ids for training.
65
+
66
+ Returns:
67
+ list[dict]: a list of dicts in Detectron2 standard format. (See
68
+ `Using Custom Datasets </tutorials/datasets.html>`_ )
69
+ """
70
+
71
+ def _convert_category_id(segment_info, meta):
72
+ if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]:
73
+ segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][
74
+ segment_info["category_id"]
75
+ ]
76
+ else:
77
+ segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][
78
+ segment_info["category_id"]
79
+ ]
80
+ return segment_info
81
+
82
+ assert os.path.exists(
83
+ gt_json
84
+ ), "Please run `python cityscapesscripts/preparation/createPanopticImgs.py` to generate label files." # noqa
85
+
86
+
87
+ with open(gt_json) as f:
88
+ json_info = json.load(f)
89
+
90
+ files = get_cityscapes_panoptic_files(image_dir, gt_dir, json_info)
91
+ ret = []
92
+ for image_file, label_file, segments_info in files:
93
+ sem_label_file = (
94
+ image_file.replace("leftImg8bit", "gtFine").split(".")[0] + "_labelTrainIds.png"
95
+ )
96
+ segments_info = [_convert_category_id(x, meta) for x in segments_info]
97
+ ret.append(
98
+ {
99
+ "file_name": image_file,
100
+ "image_id": "_".join(
101
+ os.path.splitext(os.path.basename(image_file))[0].split("_")[:3]
102
+ ),
103
+ "sem_seg_file_name": sem_label_file,
104
+ "pan_seg_file_name": label_file,
105
+ "segments_info": segments_info,
106
+ }
107
+ )
108
+ assert len(ret), f"No images found in {image_dir}!"
109
+ assert PathManager.isfile(
110
+ ret[0]["sem_seg_file_name"]
111
+ ), "Please generate labelTrainIds.png with cityscapesscripts/preparation/createTrainIdLabelImgs.py" # noqa
112
+ assert PathManager.isfile(
113
+ ret[0]["pan_seg_file_name"]
114
+ ), "Please generate panoptic annotation with python cityscapesscripts/preparation/createPanopticImgs.py" # noqa
115
+ return ret
116
+
117
+
118
+ _RAW_CITYSCAPES_PANOPTIC_SPLITS = {
119
+ "cityscapes_fine_panoptic_train": (
120
+ "cityscapes/leftImg8bit/train",
121
+ "cityscapes/gtFine/cityscapes_panoptic_train",
122
+ "cityscapes/gtFine/cityscapes_panoptic_train.json",
123
+ ),
124
+ "cityscapes_fine_panoptic_val": (
125
+ "cityscapes/leftImg8bit/val",
126
+ "cityscapes/gtFine/cityscapes_panoptic_val",
127
+ "cityscapes/gtFine/cityscapes_panoptic_val.json",
128
+ ),
129
+ # "cityscapes_fine_panoptic_test": not supported yet
130
+ }
131
+
132
+
133
+ def register_all_cityscapes_panoptic(root):
134
+ meta = {}
135
+ # The following metadata maps contiguous id from [0, #thing categories +
136
+ # #stuff categories) to their names and colors. We have to replica of the
137
+ # same name and color under "thing_*" and "stuff_*" because the current
138
+ # visualization function in D2 handles thing and class classes differently
139
+ # due to some heuristic used in Panoptic FPN. We keep the same naming to
140
+ # enable reusing existing visualization functions.
141
+ thing_classes = [k["name"] for k in CITYSCAPES_CATEGORIES]
142
+ thing_colors = [k["color"] for k in CITYSCAPES_CATEGORIES]
143
+ stuff_classes = [k["name"] for k in CITYSCAPES_CATEGORIES]
144
+ stuff_colors = [k["color"] for k in CITYSCAPES_CATEGORIES]
145
+
146
+ meta["thing_classes"] = thing_classes
147
+ meta["thing_colors"] = thing_colors
148
+ meta["stuff_classes"] = stuff_classes
149
+ meta["stuff_colors"] = stuff_colors
150
+
151
+ # There are three types of ids in cityscapes panoptic segmentation:
152
+ # (1) category id: like semantic segmentation, it is the class id for each
153
+ # pixel. Since there are some classes not used in evaluation, the category
154
+ # id is not always contiguous and thus we have two set of category ids:
155
+ # - original category id: category id in the original dataset, mainly
156
+ # used for evaluation.
157
+ # - contiguous category id: [0, #classes), in order to train the classifier
158
+ # (2) instance id: this id is used to differentiate different instances from
159
+ # the same category. For "stuff" classes, the instance id is always 0; for
160
+ # "thing" classes, the instance id starts from 1 and 0 is reserved for
161
+ # ignored instances (e.g. crowd annotation).
162
+ # (3) panoptic id: this is the compact id that encode both category and
163
+ # instance id by: category_id * 1000 + instance_id.
164
+ thing_dataset_id_to_contiguous_id = {}
165
+ stuff_dataset_id_to_contiguous_id = {}
166
+
167
+ for k in CITYSCAPES_CATEGORIES:
168
+ if k["isthing"] == 1:
169
+ thing_dataset_id_to_contiguous_id[k["id"]] = k["trainId"]
170
+ else:
171
+ stuff_dataset_id_to_contiguous_id[k["id"]] = k["trainId"]
172
+
173
+ meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id
174
+ meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id
175
+
176
+ for key, (image_dir, gt_dir, gt_json) in _RAW_CITYSCAPES_PANOPTIC_SPLITS.items():
177
+ image_dir = os.path.join(root, image_dir)
178
+ gt_dir = os.path.join(root, gt_dir)
179
+ gt_json = os.path.join(root, gt_json)
180
+
181
+ if key in DatasetCatalog.list():
182
+ DatasetCatalog.remove(key)
183
+
184
+ DatasetCatalog.register(
185
+ key, lambda x=image_dir, y=gt_dir, z=gt_json: load_cityscapes_panoptic(x, y, z, meta)
186
+ )
187
+ MetadataCatalog.get(key).set(
188
+ panoptic_root=gt_dir,
189
+ image_root=image_dir,
190
+ panoptic_json=gt_json,
191
+ gt_dir=gt_dir.replace("cityscapes_panoptic_", ""),
192
+ evaluator_type="cityscapes_panoptic_seg",
193
+ ignore_label=255,
194
+ label_divisor=1000,
195
+ **meta,
196
+ )
197
+
198
+ _root = os.getenv("DETECTRON2_DATASETS", "datasets")
199
+ register_all_cityscapes_panoptic(_root)
oneformer/data/datasets/register_coco_panoptic2instance.py ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/datasets/builtin.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+
7
+ """
8
+ This file registers pre-defined datasets at hard-coded paths, and their metadata.
9
+
10
+ We hard-code metadata for common datasets. This will enable:
11
+ 1. Consistency check when loading the datasets
12
+ 2. Use models on these standard datasets directly and run demos,
13
+ without having to download the dataset annotations
14
+
15
+ We hard-code some paths to the dataset that's assumed to
16
+ exist in "./datasets/".
17
+
18
+ Users SHOULD NOT use this file to create new dataset / metadata for new dataset.
19
+ To add new dataset, refer to the tutorial "docs/DATASETS.md".
20
+ """
21
+
22
+ import os
23
+ from detectron2.data.datasets.builtin_meta import _get_builtin_metadata
24
+ from detectron2.data.datasets.coco import register_coco_instances
25
+
26
+
27
+ _PREDEFINED_SPLITS_COCO = {
28
+ "coco_2017_val_panoptic2instance": ("coco/val2017", "coco/annotations/panoptic2instances_val2017.json"),
29
+ }
30
+
31
+
32
+ def register_panoptic2instances_coco(root):
33
+ for key, (image_root, json_file) in _PREDEFINED_SPLITS_COCO.items():
34
+ # Assume pre-defined datasets live in `./datasets`.
35
+ register_coco_instances(
36
+ key,
37
+ _get_builtin_metadata("coco"),
38
+ os.path.join(root, json_file) if "://" not in json_file else json_file,
39
+ os.path.join(root, image_root),
40
+ )
41
+
42
+
43
+ _root = os.path.expanduser(os.getenv("DETECTRON2_DATASETS", "datasets"))
44
+ register_panoptic2instances_coco(_root)
oneformer/data/datasets/register_coco_panoptic_annos_semseg.py ADDED
@@ -0,0 +1,367 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/data/datasets/register_coco_panoptic_annos_semseg.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import json
7
+ import os
8
+
9
+ from detectron2.data import DatasetCatalog, MetadataCatalog
10
+ from detectron2.data.datasets import load_sem_seg
11
+ from detectron2.data.datasets.builtin_meta import COCO_CATEGORIES
12
+ from detectron2.utils.file_io import PathManager
13
+ import contextlib
14
+ import logging
15
+ import io
16
+ from fvcore.common.timer import Timer
17
+ import pycocotools.mask as mask_util
18
+ from detectron2.structures import BoxMode
19
+
20
+
21
+ logger = logging.getLogger(__name__)
22
+
23
+
24
+ _PREDEFINED_SPLITS_COCO_PANOPTIC = {
25
+ "coco_2017_train_panoptic": (
26
+ # This is the original panoptic annotation directory
27
+ "coco/panoptic_train2017",
28
+ "coco/annotations/panoptic_train2017.json",
29
+ # This directory contains semantic annotations that are
30
+ # converted from panoptic annotations.
31
+ # It is used by PanopticFPN.
32
+ # You can use the script at detectron2/datasets/prepare_panoptic_fpn.py
33
+ # to create these directories.
34
+ "coco/panoptic_semseg_train2017",
35
+ ),
36
+ "coco_2017_val_panoptic": (
37
+ "coco/panoptic_val2017",
38
+ "coco/annotations/panoptic_val2017.json",
39
+ "coco/panoptic_semseg_val2017",
40
+ ),
41
+ }
42
+
43
+ def load_coco_instance_json(json_file, image_root, dataset_name=None):
44
+ from pycocotools.coco import COCO
45
+
46
+ timer = Timer()
47
+ json_file = PathManager.get_local_path(json_file)
48
+ with contextlib.redirect_stdout(io.StringIO()):
49
+ coco_api = COCO(json_file)
50
+ if timer.seconds() > 1:
51
+ logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds()))
52
+
53
+ id_map = None
54
+ if dataset_name is not None:
55
+ meta = MetadataCatalog.get(dataset_name)
56
+ cat_ids = sorted(coco_api.getCatIds())
57
+ cats = coco_api.loadCats(cat_ids)
58
+ # The categories in a custom json file may not be sorted.
59
+ thing_classes = [c["name"] for c in sorted(cats, key=lambda x: x["id"])]
60
+ meta.thing_classes = thing_classes
61
+
62
+ # In COCO, certain category ids are artificially removed,
63
+ # and by convention they are always ignored.
64
+ # We deal with COCO's id issue and translate
65
+ # the category ids to contiguous ids in [0, 80).
66
+
67
+ # It works by looking at the "categories" field in the json, therefore
68
+ # if users' own json also have incontiguous ids, we'll
69
+ # apply this mapping as well but print a warning.
70
+ if not (min(cat_ids) == 1 and max(cat_ids) == len(cat_ids)):
71
+ if "coco" not in dataset_name:
72
+ logger.warning(
73
+ """
74
+ Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.
75
+ """
76
+ )
77
+ id_map = {v: i for i, v in enumerate(cat_ids)}
78
+ meta.thing_dataset_id_to_contiguous_id = id_map
79
+
80
+ # sort indices for reproducible results
81
+ img_ids = sorted(coco_api.imgs.keys())
82
+ # imgs is a list of dicts, each looks something like:
83
+ # {'license': 4,
84
+ # 'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
85
+ # 'file_name': 'COCO_val2014_000000001268.jpg',
86
+ # 'height': 427,
87
+ # 'width': 640,
88
+ # 'date_captured': '2013-11-17 05:57:24',
89
+ # 'id': 1268}
90
+ imgs = coco_api.loadImgs(img_ids)
91
+ # anns is a list[list[dict]], where each dict is an annotation
92
+ # record for an object. The inner list enumerates the objects in an image
93
+ # and the outer list enumerates over images. Example of anns[0]:
94
+ # [{'segmentation': [[192.81,
95
+ # 247.09,
96
+ # ...
97
+ # 219.03,
98
+ # 249.06]],
99
+ # 'area': 1035.749,
100
+ # 'iscrowd': 0,
101
+ # 'image_id': 1268,
102
+ # 'bbox': [192.81, 224.8, 74.73, 33.43],
103
+ # 'category_id': 16,
104
+ # 'id': 42986},
105
+ # ...]
106
+ anns = [coco_api.imgToAnns[img_id] for img_id in img_ids]
107
+ total_num_valid_anns = sum([len(x) for x in anns])
108
+ total_num_anns = len(coco_api.anns)
109
+ if total_num_valid_anns < total_num_anns:
110
+ logger.warning(
111
+ f"{json_file} contains {total_num_anns} annotations, but only "
112
+ f"{total_num_valid_anns} of them match to images in the file."
113
+ )
114
+
115
+ if "minival" not in json_file:
116
+ # The popular valminusminival & minival annotations for COCO2014 contain this bug.
117
+ # However the ratio of buggy annotations there is tiny and does not affect accuracy.
118
+ # Therefore we explicitly white-list them.
119
+ ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image]
120
+ assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique!".format(
121
+ json_file
122
+ )
123
+
124
+ imgs_anns = list(zip(imgs, anns))
125
+ logger.info("Loaded {} images in COCO format from {}".format(len(imgs_anns), json_file))
126
+
127
+ dataset_dicts = {}
128
+
129
+ ann_keys = ["iscrowd", "bbox", "keypoints", "category_id"]
130
+
131
+ num_instances_without_valid_segmentation = 0
132
+
133
+ for (img_dict, anno_dict_list) in imgs_anns:
134
+ record = {}
135
+ record["file_name"] = os.path.join(image_root, img_dict["file_name"])
136
+ record["height"] = img_dict["height"]
137
+ record["width"] = img_dict["width"]
138
+ image_id = record["image_id"] = img_dict["id"]
139
+
140
+ objs = []
141
+ for anno in anno_dict_list:
142
+ # Check that the image_id in this annotation is the same as
143
+ # the image_id we're looking at.
144
+ # This fails only when the data parsing logic or the annotation file is buggy.
145
+
146
+ # The original COCO valminusminival2014 & minival2014 annotation files
147
+ # actually contains bugs that, together with certain ways of using COCO API,
148
+ # can trigger this assertion.
149
+ assert anno["image_id"] == image_id
150
+
151
+ assert anno.get("ignore", 0) == 0, '"ignore" in COCO json file is not supported.'
152
+
153
+ obj = {key: anno[key] for key in ann_keys if key in anno}
154
+ if "bbox" in obj and len(obj["bbox"]) == 0:
155
+ raise ValueError(
156
+ f"One annotation of image {image_id} contains empty 'bbox' value! "
157
+ "This json does not have valid COCO format."
158
+ )
159
+
160
+ segm = anno.get("segmentation", None)
161
+ if segm: # either list[list[float]] or dict(RLE)
162
+ if isinstance(segm, dict):
163
+ if isinstance(segm["counts"], list):
164
+ # convert to compressed RLE
165
+ segm = mask_util.frPyObjects(segm, *segm["size"])
166
+ else:
167
+ # filter out invalid polygons (< 3 points)
168
+ segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6]
169
+ if len(segm) == 0:
170
+ num_instances_without_valid_segmentation += 1
171
+ continue # ignore this instance
172
+ obj["segmentation"] = segm
173
+
174
+ keypts = anno.get("keypoints", None)
175
+ if keypts: # list[int]
176
+ for idx, v in enumerate(keypts):
177
+ if idx % 3 != 2:
178
+ # COCO's segmentation coordinates are floating points in [0, H or W],
179
+ # but keypoint coordinates are integers in [0, H-1 or W-1]
180
+ # Therefore we assume the coordinates are "pixel indices" and
181
+ # add 0.5 to convert to floating point coordinates.
182
+ keypts[idx] = v + 0.5
183
+ obj["keypoints"] = keypts
184
+
185
+ obj["bbox_mode"] = BoxMode.XYWH_ABS
186
+ if id_map:
187
+ annotation_category_id = obj["category_id"]
188
+ try:
189
+ obj["category_id"] = id_map[annotation_category_id]
190
+ except KeyError as e:
191
+ raise KeyError(
192
+ f"Encountered category_id={annotation_category_id} "
193
+ "but this id does not exist in 'categories' of the json file."
194
+ ) from e
195
+ objs.append(obj)
196
+ record["annotations"] = objs
197
+ dataset_dicts[image_id] = record
198
+
199
+ if num_instances_without_valid_segmentation > 0:
200
+ logger.warning(
201
+ "Filtered out {} instances without valid segmentation. ".format(
202
+ num_instances_without_valid_segmentation
203
+ )
204
+ + "There might be issues in your dataset generation process. Please "
205
+ "check https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html carefully"
206
+ )
207
+ return dataset_dicts
208
+
209
+ def get_metadata():
210
+ meta = {}
211
+ # The following metadata maps contiguous id from [0, #thing categories +
212
+ # #stuff categories) to their names and colors. We have to replica of the
213
+ # same name and color under "thing_*" and "stuff_*" because the current
214
+ # visualization function in D2 handles thing and class classes differently
215
+ # due to some heuristic used in Panoptic FPN. We keep the same naming to
216
+ # enable reusing existing visualization functions.
217
+ thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1]
218
+ thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1]
219
+ stuff_classes = [k["name"] for k in COCO_CATEGORIES]
220
+ stuff_colors = [k["color"] for k in COCO_CATEGORIES]
221
+
222
+ meta["thing_classes"] = thing_classes
223
+ meta["thing_colors"] = thing_colors
224
+ meta["stuff_classes"] = stuff_classes
225
+ meta["stuff_colors"] = stuff_colors
226
+
227
+ # Convert category id for training:
228
+ # category id: like semantic segmentation, it is the class id for each
229
+ # pixel. Since there are some classes not used in evaluation, the category
230
+ # id is not always contiguous and thus we have two set of category ids:
231
+ # - original category id: category id in the original dataset, mainly
232
+ # used for evaluation.
233
+ # - contiguous category id: [0, #classes), in order to train the linear
234
+ # softmax classifier.
235
+ thing_dataset_id_to_contiguous_id = {}
236
+ stuff_dataset_id_to_contiguous_id = {}
237
+
238
+ for i, cat in enumerate(COCO_CATEGORIES):
239
+ if cat["isthing"]:
240
+ thing_dataset_id_to_contiguous_id[cat["id"]] = i
241
+ # else:
242
+ # stuff_dataset_id_to_contiguous_id[cat["id"]] = i
243
+
244
+ # in order to use sem_seg evaluator
245
+ stuff_dataset_id_to_contiguous_id[cat["id"]] = i
246
+
247
+ meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id
248
+ meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id
249
+
250
+ return meta
251
+
252
+
253
+ def load_coco_panoptic_json(json_file, instances_json, instances_name, image_dir, gt_dir, semseg_dir, meta):
254
+ """
255
+ Args:
256
+ image_dir (str): path to the raw dataset. e.g., "~/coco/train2017".
257
+ gt_dir (str): path to the raw annotations. e.g., "~/coco/panoptic_train2017".
258
+ json_file (str): path to the json file. e.g., "~/coco/annotations/panoptic_train2017.json".
259
+ Returns:
260
+ list[dict]: a list of dicts in Detectron2 standard format. (See
261
+ `Using Custom Datasets </tutorials/datasets.html>`_ )
262
+ """
263
+
264
+ def _convert_category_id(segment_info, meta):
265
+ if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]:
266
+ segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][
267
+ segment_info["category_id"]
268
+ ]
269
+ segment_info["isthing"] = True
270
+ else:
271
+ segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][
272
+ segment_info["category_id"]
273
+ ]
274
+ segment_info["isthing"] = False
275
+ return segment_info
276
+
277
+ with PathManager.open(json_file) as f:
278
+ json_info = json.load(f)
279
+
280
+ instance_data_dicts = load_coco_instance_json(instances_json, image_dir.replace("panoptic_", ""), instances_name)
281
+
282
+ ret = []
283
+ for ann in json_info["annotations"]:
284
+ image_id = int(ann["image_id"])
285
+ # TODO: currently we assume image and label has the same filename but
286
+ # different extension, and images have extension ".jpg" for COCO. Need
287
+ # to make image extension a user-provided argument if we extend this
288
+ # function to support other COCO-like datasets.
289
+ image_file = os.path.join(image_dir, os.path.splitext(ann["file_name"])[0] + ".jpg")
290
+ label_file = os.path.join(gt_dir, ann["file_name"])
291
+ sem_label_file = os.path.join(semseg_dir, ann["file_name"])
292
+ segments_info = [_convert_category_id(x, meta) for x in ann["segments_info"]]
293
+ ret.append(
294
+ {
295
+ "file_name": image_file,
296
+ "image_id": image_id,
297
+ "pan_seg_file_name": label_file,
298
+ "sem_seg_file_name": sem_label_file,
299
+ "segments_info": segments_info,
300
+ "annotations": instance_data_dicts[image_id]["annotations"],
301
+ }
302
+ )
303
+ assert len(ret), f"No images found in {image_dir}!"
304
+ assert PathManager.isfile(ret[0]["file_name"]), ret[0]["file_name"]
305
+ assert PathManager.isfile(ret[0]["pan_seg_file_name"]), ret[0]["pan_seg_file_name"]
306
+ assert PathManager.isfile(ret[0]["sem_seg_file_name"]), ret[0]["sem_seg_file_name"]
307
+ return ret
308
+
309
+
310
+ def register_coco_panoptic_annos_sem_seg(
311
+ name, metadata, image_root, panoptic_root, panoptic_json, sem_seg_root, instances_json, instances_name,
312
+ ):
313
+ panoptic_name = name
314
+ delattr(MetadataCatalog.get(panoptic_name), "thing_classes")
315
+ delattr(MetadataCatalog.get(panoptic_name), "thing_colors")
316
+ MetadataCatalog.get(panoptic_name).set(
317
+ thing_classes=metadata["thing_classes"],
318
+ thing_colors=metadata["thing_colors"],
319
+ # thing_dataset_id_to_contiguous_id=metadata["thing_dataset_id_to_contiguous_id"],
320
+ )
321
+
322
+ # the name is "coco_2017_train_panoptic_with_sem_seg" and "coco_2017_val_panoptic_with_sem_seg"
323
+ semantic_name = name + "_with_sem_seg"
324
+ DatasetCatalog.register(
325
+ semantic_name,
326
+ lambda: load_coco_panoptic_json(panoptic_json, instances_json, instances_name, image_root, panoptic_root, sem_seg_root, metadata),
327
+ )
328
+ MetadataCatalog.get(semantic_name).set(
329
+ sem_seg_root=sem_seg_root,
330
+ panoptic_root=panoptic_root,
331
+ image_root=image_root,
332
+ panoptic_json=panoptic_json,
333
+ json_file=instances_json,
334
+ evaluator_type="coco_panoptic_seg",
335
+ ignore_label=255,
336
+ label_divisor=1000,
337
+ **metadata,
338
+ )
339
+
340
+
341
+ def register_all_coco_panoptic_annos_sem_seg(root):
342
+ for (
343
+ prefix,
344
+ (panoptic_root, panoptic_json, semantic_root),
345
+ ) in _PREDEFINED_SPLITS_COCO_PANOPTIC.items():
346
+
347
+ prefix_instances = prefix[: -len("_panoptic")]
348
+ instances_meta = MetadataCatalog.get(prefix_instances)
349
+ image_root, instances_json = instances_meta.image_root, instances_meta.json_file
350
+
351
+ if 'val' in instances_json:
352
+ instances_json = instances_json.replace('instances_', 'panoptic2instances_')
353
+
354
+ register_coco_panoptic_annos_sem_seg(
355
+ prefix,
356
+ get_metadata(),
357
+ image_root,
358
+ os.path.join(root, panoptic_root),
359
+ os.path.join(root, panoptic_json),
360
+ os.path.join(root, semantic_root),
361
+ instances_json,
362
+ prefix_instances,
363
+ )
364
+
365
+
366
+ _root = os.getenv("DETECTRON2_DATASETS", "datasets")
367
+ register_all_coco_panoptic_annos_sem_seg(_root)
oneformer/data/tokenizer.py ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -------------------------------------------------------------------------
2
+ # MIT License
3
+ #
4
+ # Copyright (c) 2021 OpenAI
5
+ #
6
+ # Permission is hereby granted, free of charge, to any person obtaining a copy
7
+ # of this software and associated documentation files (the "Software"), to deal
8
+ # in the Software without restriction, including without limitation the rights
9
+ # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10
+ # copies of the Software, and to permit persons to whom the Software is
11
+ # furnished to do so, subject to the following conditions:
12
+ #
13
+ # The above copyright notice and this permission notice shall be included in all
14
+ # copies or substantial portions of the Software.
15
+ #
16
+ # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17
+ # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18
+ # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19
+ # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20
+ # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21
+ # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22
+ # SOFTWARE.
23
+ #
24
+ # Modified by Jiarui Xu
25
+ # -------------------------------------------------------------------------
26
+
27
+ import gzip
28
+ import html
29
+ import os
30
+ from functools import lru_cache
31
+
32
+ import ftfy
33
+ import regex as re
34
+ import torch
35
+
36
+
37
+ @lru_cache()
38
+ def default_bpe():
39
+ return os.path.join(os.path.dirname(os.path.abspath(__file__)), 'bpe_simple_vocab_16e6.txt.gz')
40
+
41
+
42
+ @lru_cache()
43
+ def bytes_to_unicode():
44
+ """Returns list of utf-8 byte and a corresponding list of unicode strings.
45
+
46
+ The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
47
+ if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for decent
48
+ coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup tables
49
+ between utf-8 bytes and unicode strings. And avoids mapping to whitespace/control characters the bpe code barfs on.
50
+ """
51
+ bs = list(range(ord('!'), ord('~') + 1)) + list(range(ord('¡'), ord('¬') + 1)) + list(range(ord('®'), ord('ÿ') + 1))
52
+ cs = bs[:]
53
+ n = 0
54
+ for b in range(2**8):
55
+ if b not in bs:
56
+ bs.append(b)
57
+ cs.append(2**8 + n)
58
+ n += 1
59
+ cs = [chr(n) for n in cs]
60
+ return dict(zip(bs, cs))
61
+
62
+
63
+ def get_pairs(word):
64
+ """Return set of symbol pairs in a word.
65
+
66
+ Word is represented as tuple of symbols (symbols being variable-length strings).
67
+ """
68
+ pairs = set()
69
+ prev_char = word[0]
70
+ for char in word[1:]:
71
+ pairs.add((prev_char, char))
72
+ prev_char = char
73
+ return pairs
74
+
75
+
76
+ def basic_clean(text):
77
+ text = ftfy.fix_text(text)
78
+ text = html.unescape(html.unescape(text))
79
+ return text.strip()
80
+
81
+
82
+ def whitespace_clean(text):
83
+ text = re.sub(r'\s+', ' ', text)
84
+ text = text.strip()
85
+ return text
86
+
87
+ class Tokenize:
88
+
89
+ def __init__(self, tokenizer, max_seq_len=77, truncate=True):
90
+ self.tokenizer = tokenizer
91
+ self.max_seq_len = max_seq_len
92
+ self.truncate = truncate
93
+
94
+ def __call__(self, texts):
95
+ expanded_dim = False
96
+ if isinstance(texts, str):
97
+ texts = [texts]
98
+ expanded_dim = True
99
+
100
+ sot_token = self.tokenizer.encoder['<|startoftext|>']
101
+ eot_token = self.tokenizer.encoder['<|endoftext|>']
102
+ all_tokens = [[sot_token] + self.tokenizer.encode(text) + [eot_token] for text in texts]
103
+ result = torch.zeros(len(all_tokens), self.max_seq_len, dtype=torch.long)
104
+
105
+ for i, tokens in enumerate(all_tokens):
106
+ if len(tokens) > self.max_seq_len:
107
+ if self.truncate:
108
+ tokens = tokens[:self.max_seq_len]
109
+ tokens[-1] = eot_token
110
+ else:
111
+ raise RuntimeError(f'Input {texts[i]} is too long for context length {self.max_seq_len}')
112
+ result[i, :len(tokens)] = torch.tensor(tokens)
113
+
114
+ if expanded_dim:
115
+ return result[0]
116
+
117
+ return result
118
+
119
+
120
+ class SimpleTokenizer(object):
121
+
122
+ def __init__(self, bpe_path: str = default_bpe()):
123
+ self.byte_encoder = bytes_to_unicode()
124
+ self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
125
+ merges = gzip.open(bpe_path).read().decode('utf-8').split('\n')
126
+ merges = merges[1:49152 - 256 - 2 + 1]
127
+ merges = [tuple(merge.split()) for merge in merges]
128
+ vocab = list(bytes_to_unicode().values())
129
+ vocab = vocab + [v + '</w>' for v in vocab]
130
+ for merge in merges:
131
+ vocab.append(''.join(merge))
132
+ vocab.extend(['<|startoftext|>', '<|endoftext|>'])
133
+ self.encoder = dict(zip(vocab, range(len(vocab))))
134
+ self.decoder = {v: k for k, v in self.encoder.items()}
135
+ self.bpe_ranks = dict(zip(merges, range(len(merges))))
136
+ self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
137
+ self.pat = re.compile(
138
+ r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
139
+ re.IGNORECASE)
140
+
141
+ def bpe(self, token):
142
+ if token in self.cache:
143
+ return self.cache[token]
144
+ word = tuple(token[:-1]) + (token[-1] + '</w>', )
145
+ pairs = get_pairs(word)
146
+
147
+ if not pairs:
148
+ return token + '</w>'
149
+
150
+ while True:
151
+ bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float('inf')))
152
+ if bigram not in self.bpe_ranks:
153
+ break
154
+ first, second = bigram
155
+ new_word = []
156
+ i = 0
157
+ while i < len(word):
158
+ try:
159
+ j = word.index(first, i)
160
+ new_word.extend(word[i:j])
161
+ i = j
162
+ except: # noqa: E722
163
+ new_word.extend(word[i:])
164
+ break
165
+
166
+ if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
167
+ new_word.append(first + second)
168
+ i += 2
169
+ else:
170
+ new_word.append(word[i])
171
+ i += 1
172
+ new_word = tuple(new_word)
173
+ word = new_word
174
+ if len(word) == 1:
175
+ break
176
+ else:
177
+ pairs = get_pairs(word)
178
+ word = ' '.join(word)
179
+ self.cache[token] = word
180
+ return word
181
+
182
+ def encode(self, text):
183
+ bpe_tokens = []
184
+ text = whitespace_clean(basic_clean(text)).lower()
185
+ for token in re.findall(self.pat, text):
186
+ token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
187
+ bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
188
+ return bpe_tokens
189
+
190
+ def decode(self, tokens):
191
+ text = ''.join([self.decoder[token] for token in tokens])
192
+ text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors='replace').replace('</w>', ' ')
193
+ return text
oneformer/evaluation/__init__.py ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ from .detection_coco_evaluator import *
2
+ from .coco_evaluator import *
3
+ from .cityscapes_evaluation import CityscapesInstanceEvaluator
oneformer/evaluation/cityscapes_evaluation.py ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/cityscapes_evaluation.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import glob
7
+ import logging
8
+ import numpy as np
9
+ import os
10
+ import tempfile
11
+ from collections import OrderedDict
12
+ import torch
13
+ from PIL import Image
14
+
15
+ from detectron2.data import MetadataCatalog
16
+ from detectron2.utils import comm
17
+ from detectron2.utils.file_io import PathManager
18
+
19
+ from .evaluator import DatasetEvaluator
20
+
21
+
22
+ class CityscapesEvaluator(DatasetEvaluator):
23
+ """
24
+ Base class for evaluation using cityscapes API.
25
+ """
26
+
27
+ def __init__(self, dataset_name):
28
+ """
29
+ Args:
30
+ dataset_name (str): the name of the dataset.
31
+ It must have the following metadata associated with it:
32
+ "thing_classes", "gt_dir".
33
+ """
34
+ self._metadata = MetadataCatalog.get(dataset_name)
35
+ self._cpu_device = torch.device("cpu")
36
+ self._logger = logging.getLogger(__name__)
37
+
38
+ def reset(self):
39
+ self._working_dir = tempfile.TemporaryDirectory(prefix="cityscapes_eval_")
40
+ self._temp_dir = self._working_dir.name
41
+ # All workers will write to the same results directory
42
+ # TODO this does not work in distributed training
43
+ assert (
44
+ comm.get_local_size() == comm.get_world_size()
45
+ ), "CityscapesEvaluator currently do not work with multiple machines."
46
+ self._temp_dir = comm.all_gather(self._temp_dir)[0]
47
+ if self._temp_dir != self._working_dir.name:
48
+ self._working_dir.cleanup()
49
+ self._logger.info(
50
+ "Writing cityscapes results to temporary directory {} ...".format(self._temp_dir)
51
+ )
52
+
53
+
54
+ class CityscapesInstanceEvaluator(CityscapesEvaluator):
55
+ """
56
+ Evaluate instance segmentation results on cityscapes dataset using cityscapes API.
57
+
58
+ Note:
59
+ * It does not work in multi-machine distributed training.
60
+ * It contains a synchronization, therefore has to be used on all ranks.
61
+ * Only the main process runs evaluation.
62
+ """
63
+
64
+ def process(self, inputs, outputs):
65
+ from cityscapesscripts.helpers.labels import name2label
66
+
67
+ for input, output in zip(inputs, outputs):
68
+ file_name = input["file_name"]
69
+ basename = os.path.splitext(os.path.basename(file_name))[0]
70
+ pred_txt = os.path.join(self._temp_dir, basename + "_pred.txt")
71
+
72
+ if "instances" in output:
73
+ output = output["instances"].to(self._cpu_device)
74
+ num_instances = len(output)
75
+ with open(pred_txt, "w") as fout:
76
+ for i in range(num_instances):
77
+ pred_class = output.pred_classes[i]
78
+ classes = self._metadata.stuff_classes[pred_class]
79
+ class_id = name2label[classes].id
80
+ score = output.scores[i]
81
+ mask = output.pred_masks[i].numpy().astype("uint8")
82
+ png_filename = os.path.join(
83
+ self._temp_dir, basename + "_{}_{}.png".format(i, classes)
84
+ )
85
+
86
+ Image.fromarray(mask * 255).save(png_filename)
87
+ fout.write(
88
+ "{} {} {}\n".format(os.path.basename(png_filename), class_id, score)
89
+ )
90
+ else:
91
+ # Cityscapes requires a prediction file for every ground truth image.
92
+ with open(pred_txt, "w") as fout:
93
+ pass
94
+
95
+ def evaluate(self):
96
+ """
97
+ Returns:
98
+ dict: has a key "segm", whose value is a dict of "AP" and "AP50".
99
+ """
100
+ comm.synchronize()
101
+ if comm.get_rank() > 0:
102
+ return
103
+ import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as cityscapes_eval
104
+
105
+ self._logger.info("Evaluating results under {} ...".format(self._temp_dir))
106
+
107
+ # set some global states in cityscapes evaluation API, before evaluating
108
+ cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir)
109
+ cityscapes_eval.args.predictionWalk = None
110
+ cityscapes_eval.args.JSONOutput = False
111
+ cityscapes_eval.args.colorized = False
112
+ cityscapes_eval.args.gtInstancesFile = os.path.join(self._temp_dir, "gtInstances.json")
113
+
114
+ # These lines are adopted from
115
+ # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalInstanceLevelSemanticLabeling.py # noqa
116
+ gt_dir = PathManager.get_local_path(self._metadata.gt_dir)
117
+ groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_instanceIds.png"))
118
+ assert len(
119
+ groundTruthImgList
120
+ ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format(
121
+ cityscapes_eval.args.groundTruthSearch
122
+ )
123
+ predictionImgList = []
124
+ for gt in groundTruthImgList:
125
+ predictionImgList.append(cityscapes_eval.getPrediction(gt, cityscapes_eval.args))
126
+ results = cityscapes_eval.evaluateImgLists(
127
+ predictionImgList, groundTruthImgList, cityscapes_eval.args
128
+ )["averages"]
129
+
130
+ ret = OrderedDict()
131
+ ret["segm"] = {"AP": results["allAp"] * 100, "AP50": results["allAp50%"] * 100}
132
+ self._working_dir.cleanup()
133
+ return ret
134
+
135
+
136
+ class CityscapesSemSegEvaluator(CityscapesEvaluator):
137
+ """
138
+ Evaluate semantic segmentation results on cityscapes dataset using cityscapes API.
139
+
140
+ Note:
141
+ * It does not work in multi-machine distributed training.
142
+ * It contains a synchronization, therefore has to be used on all ranks.
143
+ * Only the main process runs evaluation.
144
+ """
145
+
146
+ def process(self, inputs, outputs):
147
+ from cityscapesscripts.helpers.labels import trainId2label
148
+
149
+ for input, output in zip(inputs, outputs):
150
+ file_name = input["file_name"]
151
+ basename = os.path.splitext(os.path.basename(file_name))[0]
152
+ pred_filename = os.path.join(self._temp_dir, basename + "_pred.png")
153
+
154
+ output = output["sem_seg"].argmax(dim=0).to(self._cpu_device).numpy()
155
+ pred = 255 * np.ones(output.shape, dtype=np.uint8)
156
+ for train_id, label in trainId2label.items():
157
+ if label.ignoreInEval:
158
+ continue
159
+ pred[output == train_id] = label.id
160
+ Image.fromarray(pred).save(pred_filename)
161
+
162
+ def evaluate(self):
163
+ comm.synchronize()
164
+ if comm.get_rank() > 0:
165
+ return
166
+ # Load the Cityscapes eval script *after* setting the required env var,
167
+ # since the script reads CITYSCAPES_DATASET into global variables at load time.
168
+ import cityscapesscripts.evaluation.evalPixelLevelSemanticLabeling as cityscapes_eval
169
+
170
+ self._logger.info("Evaluating results under {} ...".format(self._temp_dir))
171
+
172
+ # set some global states in cityscapes evaluation API, before evaluating
173
+ cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir)
174
+ cityscapes_eval.args.predictionWalk = None
175
+ cityscapes_eval.args.JSONOutput = False
176
+ cityscapes_eval.args.colorized = False
177
+
178
+ # These lines are adopted from
179
+ # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalPixelLevelSemanticLabeling.py # noqa
180
+ gt_dir = PathManager.get_local_path(self._metadata.gt_dir)
181
+ groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_labelIds.png"))
182
+ assert len(
183
+ groundTruthImgList
184
+ ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format(
185
+ cityscapes_eval.args.groundTruthSearch
186
+ )
187
+ predictionImgList = []
188
+ for gt in groundTruthImgList:
189
+ predictionImgList.append(cityscapes_eval.getPrediction(cityscapes_eval.args, gt))
190
+ results = cityscapes_eval.evaluateImgLists(
191
+ predictionImgList, groundTruthImgList, cityscapes_eval.args
192
+ )
193
+ ret = OrderedDict()
194
+ ret["sem_seg"] = {
195
+ "IoU": 100.0 * results["averageScoreClasses"],
196
+ "iIoU": 100.0 * results["averageScoreInstClasses"],
197
+ "IoU_sup": 100.0 * results["averageScoreCategories"],
198
+ "iIoU_sup": 100.0 * results["averageScoreInstCategories"],
199
+ }
200
+ self._working_dir.cleanup()
201
+ return ret
oneformer/evaluation/coco_evaluator.py ADDED
@@ -0,0 +1,563 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/coco_evaluation.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import contextlib
7
+ import copy
8
+ import io
9
+ import itertools
10
+ import json
11
+ import logging
12
+ import numpy as np
13
+ import os
14
+ import pickle
15
+ from collections import OrderedDict
16
+ import pycocotools.mask as mask_util
17
+ import torch
18
+ from pycocotools.coco import COCO
19
+ from pycocotools.cocoeval import COCOeval
20
+ from tabulate import tabulate
21
+
22
+ import detectron2.utils.comm as comm
23
+ from detectron2.config import CfgNode
24
+ from detectron2.data import MetadataCatalog
25
+ from detectron2.data.datasets.coco import convert_to_coco_json
26
+ from detectron2.structures import Boxes, BoxMode, pairwise_iou
27
+ from detectron2.utils.file_io import PathManager
28
+ from detectron2.utils.logger import create_small_table
29
+
30
+ from .evaluator import DatasetEvaluator
31
+
32
+ try:
33
+ from detectron2.evaluation.fast_eval_api import COCOeval_opt
34
+ except ImportError:
35
+ COCOeval_opt = COCOeval
36
+
37
+
38
+ class COCOEvaluator(DatasetEvaluator):
39
+ """
40
+ Evaluate AP for instance detection/segmentation, AP
41
+ for keypoint detection outputs using COCO's metrics.
42
+ See http://cocodataset.org/#detection-eval and
43
+ http://cocodataset.org/#keypoints-eval to understand its metrics.
44
+ The metrics range from 0 to 100 (instead of 0 to 1), where a -1 or NaN means
45
+ the metric cannot be computed (e.g. due to no predictions made).
46
+
47
+ In addition to COCO, this evaluator is able to support any bounding box detection,
48
+ instance segmentation, or keypoint detection dataset.
49
+ """
50
+
51
+ def __init__(
52
+ self,
53
+ dataset_name,
54
+ tasks=None,
55
+ distributed=True,
56
+ output_dir=None,
57
+ *,
58
+ max_dets_per_image=None,
59
+ use_fast_impl=True,
60
+ kpt_oks_sigmas=(),
61
+ allow_cached_coco=True,
62
+ ):
63
+ """
64
+ Args:
65
+ dataset_name (str): name of the dataset to be evaluated.
66
+ It must have either the following corresponding metadata:
67
+
68
+ "json_file": the path to the COCO format annotation
69
+
70
+ Or it must be in detectron2's standard dataset format
71
+ so it can be converted to COCO format automatically.
72
+ tasks (tuple[str]): tasks that can be evaluated under the given
73
+ configuration. A task is one of "bbox", "segm", "keypoints".
74
+ By default, will infer this automatically from predictions.
75
+ distributed (True): if True, will collect results from all ranks and run evaluation
76
+ in the main process.
77
+ Otherwise, will only evaluate the results in the current process.
78
+ output_dir (str): optional, an output directory to dump all
79
+ results predicted on the dataset. The dump contains two files:
80
+
81
+ 1. "instances_predictions.pth" a file that can be loaded with `torch.load` and
82
+ contains all the results in the format they are produced by the model.
83
+ 2. "coco_instances_results.json" a json file in COCO's result format.
84
+ max_dets_per_image (int): limit on the maximum number of detections per image.
85
+ By default in COCO, this limit is to 100, but this can be customized
86
+ to be greater, as is needed in evaluation metrics AP fixed and AP pool
87
+ (see https://arxiv.org/pdf/2102.01066.pdf)
88
+ This doesn't affect keypoint evaluation.
89
+ use_fast_impl (bool): use a fast but **unofficial** implementation to compute AP.
90
+ Although the results should be very close to the official implementation in COCO
91
+ API, it is still recommended to compute results with the official API for use in
92
+ papers. The faster implementation also uses more RAM.
93
+ kpt_oks_sigmas (list[float]): The sigmas used to calculate keypoint OKS.
94
+ See http://cocodataset.org/#keypoints-eval
95
+ When empty, it will use the defaults in COCO.
96
+ Otherwise it should be the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS.
97
+ allow_cached_coco (bool): Whether to use cached coco json from previous validation
98
+ runs. You should set this to False if you need to use different validation data.
99
+ Defaults to True.
100
+ """
101
+ self._logger = logging.getLogger(__name__)
102
+ self._distributed = distributed
103
+ self._output_dir = output_dir
104
+
105
+ if use_fast_impl and (COCOeval_opt is COCOeval):
106
+ self._logger.info("Fast COCO eval is not built. Falling back to official COCO eval.")
107
+ use_fast_impl = False
108
+ self._use_fast_impl = use_fast_impl
109
+
110
+ # COCOeval requires the limit on the number of detections per image (maxDets) to be a list
111
+ # with at least 3 elements. The default maxDets in COCOeval is [1, 10, 100], in which the
112
+ # 3rd element (100) is used as the limit on the number of detections per image when
113
+ # evaluating AP. COCOEvaluator expects an integer for max_dets_per_image, so for COCOeval,
114
+ # we reformat max_dets_per_image into [1, 10, max_dets_per_image], based on the defaults.
115
+ if max_dets_per_image is None:
116
+ max_dets_per_image = [1, 10, 100]
117
+ else:
118
+ max_dets_per_image = [1, 10, max_dets_per_image]
119
+ self._max_dets_per_image = max_dets_per_image
120
+
121
+ if tasks is not None and isinstance(tasks, CfgNode):
122
+ kpt_oks_sigmas = (
123
+ tasks.TEST.KEYPOINT_OKS_SIGMAS if not kpt_oks_sigmas else kpt_oks_sigmas
124
+ )
125
+ self._logger.warn(
126
+ "COCO Evaluator instantiated using config, this is deprecated behavior."
127
+ " Please pass in explicit arguments instead."
128
+ )
129
+ self._tasks = None # Infering it from predictions should be better
130
+ else:
131
+ self._tasks = tasks
132
+
133
+ self._cpu_device = torch.device("cpu")
134
+
135
+ self._metadata = MetadataCatalog.get(dataset_name)
136
+ if not hasattr(self._metadata, "json_file"):
137
+ if output_dir is None:
138
+ raise ValueError(
139
+ "output_dir must be provided to COCOEvaluator "
140
+ "for datasets not in COCO format."
141
+ )
142
+ self._logger.info(f"Trying to convert '{dataset_name}' to COCO format ...")
143
+
144
+ cache_path = os.path.join(output_dir, f"{dataset_name}_coco_format.json")
145
+ self._metadata.json_file = cache_path
146
+ convert_to_coco_json(dataset_name, cache_path, allow_cached=allow_cached_coco)
147
+
148
+ json_file = PathManager.get_local_path(self._metadata.json_file)
149
+ with contextlib.redirect_stdout(io.StringIO()):
150
+ self._coco_api = COCO(json_file)
151
+
152
+ # Test set json files do not contain annotations (evaluation must be
153
+ # performed using the COCO evaluation server).
154
+ self._do_evaluation = "annotations" in self._coco_api.dataset
155
+ if self._do_evaluation:
156
+ self._kpt_oks_sigmas = kpt_oks_sigmas
157
+
158
+ def reset(self):
159
+ self._predictions = []
160
+
161
+ def process(self, inputs, outputs):
162
+ """
163
+ Args:
164
+ inputs: the inputs to a COCO model (e.g., GeneralizedRCNN).
165
+ It is a list of dict. Each dict corresponds to an image and
166
+ contains keys like "height", "width", "file_name", "image_id".
167
+ outputs: the outputs of a COCO model. It is a list of dicts with key
168
+ "instances" that contains :class:`Instances`.
169
+ """
170
+ for input, output in zip(inputs, outputs):
171
+ prediction = {"image_id": input["image_id"]}
172
+
173
+ if "instances" in output:
174
+ instances = output["instances"].to(self._cpu_device)
175
+ prediction["instances"] = instances_to_coco_json(instances, input["image_id"])
176
+ if len(prediction) > 1:
177
+ self._predictions.append(prediction)
178
+
179
+ def evaluate(self, img_ids=None):
180
+ """
181
+ Args:
182
+ img_ids: a list of image IDs to evaluate on. Default to None for the whole dataset
183
+ """
184
+ if self._distributed:
185
+ comm.synchronize()
186
+ predictions = comm.gather(self._predictions, dst=0)
187
+ predictions = list(itertools.chain(*predictions))
188
+
189
+ if not comm.is_main_process():
190
+ return {}
191
+ else:
192
+ predictions = self._predictions
193
+
194
+ if len(predictions) == 0:
195
+ self._logger.warning("[COCOEvaluator] Did not receive valid predictions.")
196
+ return {}
197
+
198
+ if self._output_dir:
199
+ PathManager.mkdirs(self._output_dir)
200
+ file_path = os.path.join(self._output_dir, "instances_predictions.pth")
201
+ with PathManager.open(file_path, "wb") as f:
202
+ torch.save(predictions, f)
203
+
204
+ self._results = OrderedDict()
205
+ if "instances" in predictions[0]:
206
+ self._eval_predictions(predictions, img_ids=img_ids)
207
+ # Copy so the caller can do whatever with results
208
+ return copy.deepcopy(self._results)
209
+
210
+ def _tasks_from_predictions(self, predictions):
211
+ """
212
+ Get COCO API "tasks" (i.e. iou_type) from COCO-format predictions.
213
+ """
214
+ for pred in predictions:
215
+ if "segmentation" in pred:
216
+ tasks = {"segm"}
217
+ if "keypoints" in pred:
218
+ tasks.add("keypoints")
219
+ return sorted(tasks)
220
+
221
+ def _eval_predictions(self, predictions, img_ids=None):
222
+ """
223
+ Evaluate predictions. Fill self._results with the metrics of the tasks.
224
+ """
225
+ self._logger.info("Preparing results for COCO format ...")
226
+ coco_results = list(itertools.chain(*[x["instances"] for x in predictions]))
227
+ tasks = self._tasks or self._tasks_from_predictions(coco_results)
228
+
229
+ # unmap the category ids for COCO
230
+ if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
231
+ dataset_id_to_contiguous_id = self._metadata.thing_dataset_id_to_contiguous_id
232
+ all_contiguous_ids = list(dataset_id_to_contiguous_id.values())
233
+ num_classes = len(all_contiguous_ids)
234
+ assert min(all_contiguous_ids) == 0 and max(all_contiguous_ids) == num_classes - 1
235
+
236
+ reverse_id_mapping = {v: k for k, v in dataset_id_to_contiguous_id.items()}
237
+ for result in coco_results:
238
+ category_id = result["category_id"]
239
+ assert category_id < num_classes, (
240
+ f"A prediction has class={category_id}, "
241
+ f"but the dataset only has {num_classes} classes and "
242
+ f"predicted class id should be in [0, {num_classes - 1}]."
243
+ )
244
+ result["category_id"] = reverse_id_mapping[category_id]
245
+
246
+ if self._output_dir:
247
+ file_path = os.path.join(self._output_dir, "coco_instances_results.json")
248
+ self._logger.info("Saving results to {}".format(file_path))
249
+ with PathManager.open(file_path, "w") as f:
250
+ f.write(json.dumps(coco_results))
251
+ f.flush()
252
+
253
+ if not self._do_evaluation:
254
+ self._logger.info("Annotations are not available for evaluation.")
255
+ return
256
+
257
+ self._logger.info(
258
+ "Evaluating predictions with {} COCO API...".format(
259
+ "unofficial" if self._use_fast_impl else "official"
260
+ )
261
+ )
262
+ for task in sorted(tasks):
263
+ assert task in {"segm", "keypoints"}, f"Got unknown task: {task}!"
264
+ coco_eval = (
265
+ _evaluate_predictions_on_coco(
266
+ self._coco_api,
267
+ coco_results,
268
+ task,
269
+ kpt_oks_sigmas=self._kpt_oks_sigmas,
270
+ use_fast_impl=self._use_fast_impl,
271
+ img_ids=img_ids,
272
+ max_dets_per_image=self._max_dets_per_image,
273
+ )
274
+ if len(coco_results) > 0
275
+ else None # cocoapi does not handle empty results very well
276
+ )
277
+
278
+ res = self._derive_coco_results(
279
+ coco_eval, task, class_names=self._metadata.get("thing_classes")
280
+ )
281
+ self._results[task] = res
282
+
283
+ def _derive_coco_results(self, coco_eval, iou_type, class_names=None):
284
+ """
285
+ Derive the desired score numbers from summarized COCOeval.
286
+
287
+ Args:
288
+ coco_eval (None or COCOEval): None represents no predictions from model.
289
+ iou_type (str):
290
+ class_names (None or list[str]): if provided, will use it to predict
291
+ per-category AP.
292
+
293
+ Returns:
294
+ a dict of {metric name: score}
295
+ """
296
+
297
+ metrics = {
298
+ "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl"],
299
+ "keypoints": ["AP", "AP50", "AP75", "APm", "APl"],
300
+ }[iou_type]
301
+
302
+ if coco_eval is None:
303
+ self._logger.warn("No predictions from the model!")
304
+ return {metric: float("nan") for metric in metrics}
305
+
306
+ # the standard metrics
307
+ results = {
308
+ metric: float(coco_eval.stats[idx] * 100 if coco_eval.stats[idx] >= 0 else "nan")
309
+ for idx, metric in enumerate(metrics)
310
+ }
311
+ self._logger.info(
312
+ "Evaluation results for {}: \n".format(iou_type) + create_small_table(results)
313
+ )
314
+ if not np.isfinite(sum(results.values())):
315
+ self._logger.info("Some metrics cannot be computed and is shown as NaN.")
316
+
317
+ if class_names is None or len(class_names) <= 1:
318
+ return results
319
+ # Compute per-category AP
320
+ # from https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L222-L252 # noqa
321
+ precisions = coco_eval.eval["precision"]
322
+ # precision has dims (iou, recall, cls, area range, max dets)
323
+ assert len(class_names) == precisions.shape[2]
324
+
325
+ results_per_category = []
326
+ for idx, name in enumerate(class_names):
327
+ # area range index 0: all area ranges
328
+ # max dets index -1: typically 100 per image
329
+ precision = precisions[:, :, idx, 0, -1]
330
+ precision = precision[precision > -1]
331
+ ap = np.mean(precision) if precision.size else float("nan")
332
+ results_per_category.append(("{}".format(name), float(ap * 100)))
333
+
334
+ # tabulate it
335
+ N_COLS = min(6, len(results_per_category) * 2)
336
+ results_flatten = list(itertools.chain(*results_per_category))
337
+ results_2d = itertools.zip_longest(*[results_flatten[i::N_COLS] for i in range(N_COLS)])
338
+ table = tabulate(
339
+ results_2d,
340
+ tablefmt="pipe",
341
+ floatfmt=".3f",
342
+ headers=["category", "AP"] * (N_COLS // 2),
343
+ numalign="left",
344
+ )
345
+ self._logger.info("Per-category {} AP: \n".format(iou_type) + table)
346
+
347
+ results.update({"AP-" + name: ap for name, ap in results_per_category})
348
+ return results
349
+
350
+
351
+ def instances_to_coco_json(instances, img_id):
352
+ """
353
+ Dump an "Instances" object to a COCO-format json that's used for evaluation.
354
+
355
+ Args:
356
+ instances (Instances):
357
+ img_id (int): the image id
358
+
359
+ Returns:
360
+ list[dict]: list of json annotations in COCO format.
361
+ """
362
+ num_instance = len(instances)
363
+ if num_instance == 0:
364
+ return []
365
+
366
+ scores = instances.scores.tolist()
367
+ classes = instances.pred_classes.tolist()
368
+
369
+ has_mask = instances.has("pred_masks")
370
+ if has_mask:
371
+ # use RLE to encode the masks, because they are too large and takes memory
372
+ # since this evaluator stores outputs of the entire dataset
373
+ rles = [
374
+ mask_util.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0]
375
+ for mask in instances.pred_masks
376
+ ]
377
+ for rle in rles:
378
+ # "counts" is an array encoded by mask_util as a byte-stream. Python3's
379
+ # json writer which always produces strings cannot serialize a bytestream
380
+ # unless you decode it. Thankfully, utf-8 works out (which is also what
381
+ # the pycocotools/_mask.pyx does).
382
+ rle["counts"] = rle["counts"].decode("utf-8")
383
+
384
+ has_keypoints = instances.has("pred_keypoints")
385
+ if has_keypoints:
386
+ keypoints = instances.pred_keypoints
387
+
388
+ results = []
389
+ for k in range(num_instance):
390
+ result = {
391
+ "image_id": img_id,
392
+ "category_id": classes[k],
393
+ "score": scores[k],
394
+ }
395
+ if has_mask:
396
+ result["segmentation"] = rles[k]
397
+ if has_keypoints:
398
+ # In COCO annotations,
399
+ # keypoints coordinates are pixel indices.
400
+ # However our predictions are floating point coordinates.
401
+ # Therefore we subtract 0.5 to be consistent with the annotation format.
402
+ # This is the inverse of data loading logic in `datasets/coco.py`.
403
+ keypoints[k][:, :2] -= 0.5
404
+ result["keypoints"] = keypoints[k].flatten().tolist()
405
+ results.append(result)
406
+ return results
407
+
408
+ def _evaluate_predictions_on_coco(
409
+ coco_gt,
410
+ coco_results,
411
+ iou_type,
412
+ kpt_oks_sigmas=None,
413
+ use_fast_impl=True,
414
+ img_ids=None,
415
+ max_dets_per_image=None,
416
+ ):
417
+ """
418
+ Evaluate the coco results using COCOEval API.
419
+ """
420
+ assert len(coco_results) > 0
421
+
422
+ if iou_type == "segm":
423
+ coco_results = copy.deepcopy(coco_results)
424
+ # When evaluating mask AP, if the results contain bbox, cocoapi will
425
+ # use the box area as the area of the instance, instead of the mask area.
426
+ # This leads to a different definition of small/medium/large.
427
+ # We remove the bbox field to let mask AP use mask area.
428
+ for c in coco_results:
429
+ c.pop("bbox", None)
430
+
431
+ coco_dt = coco_gt.loadRes(coco_results)
432
+ coco_eval = (COCOeval_opt if use_fast_impl else COCOeval)(coco_gt, coco_dt, iou_type)
433
+ # For COCO, the default max_dets_per_image is [1, 10, 100].
434
+ if max_dets_per_image is None:
435
+ max_dets_per_image = [1, 10, 100] # Default from COCOEval
436
+ else:
437
+ assert (
438
+ len(max_dets_per_image) >= 3
439
+ ), "COCOeval requires maxDets (and max_dets_per_image) to have length at least 3"
440
+ # In the case that user supplies a custom input for max_dets_per_image,
441
+ # apply COCOevalMaxDets to evaluate AP with the custom input.
442
+ if max_dets_per_image[2] != 100:
443
+ coco_eval = COCOevalMaxDets(coco_gt, coco_dt, iou_type)
444
+ if iou_type != "keypoints":
445
+ coco_eval.params.maxDets = max_dets_per_image
446
+
447
+ if img_ids is not None:
448
+ coco_eval.params.imgIds = img_ids
449
+
450
+ if iou_type == "keypoints":
451
+ # Use the COCO default keypoint OKS sigmas unless overrides are specified
452
+ if kpt_oks_sigmas:
453
+ assert hasattr(coco_eval.params, "kpt_oks_sigmas"), "pycocotools is too old!"
454
+ coco_eval.params.kpt_oks_sigmas = np.array(kpt_oks_sigmas)
455
+ # COCOAPI requires every detection and every gt to have keypoints, so
456
+ # we just take the first entry from both
457
+ num_keypoints_dt = len(coco_results[0]["keypoints"]) // 3
458
+ num_keypoints_gt = len(next(iter(coco_gt.anns.values()))["keypoints"]) // 3
459
+ num_keypoints_oks = len(coco_eval.params.kpt_oks_sigmas)
460
+ assert num_keypoints_oks == num_keypoints_dt == num_keypoints_gt, (
461
+ f"[COCOEvaluator] Prediction contain {num_keypoints_dt} keypoints. "
462
+ f"Ground truth contains {num_keypoints_gt} keypoints. "
463
+ f"The length of cfg.TEST.KEYPOINT_OKS_SIGMAS is {num_keypoints_oks}. "
464
+ "They have to agree with each other. For meaning of OKS, please refer to "
465
+ "http://cocodataset.org/#keypoints-eval."
466
+ )
467
+
468
+ coco_eval.evaluate()
469
+ coco_eval.accumulate()
470
+ coco_eval.summarize()
471
+
472
+ return coco_eval
473
+
474
+
475
+ class COCOevalMaxDets(COCOeval):
476
+ """
477
+ Modified version of COCOeval for evaluating AP with a custom
478
+ maxDets (by default for COCO, maxDets is 100)
479
+ """
480
+
481
+ def summarize(self):
482
+ """
483
+ Compute and display summary metrics for evaluation results given
484
+ a custom value for max_dets_per_image
485
+ """
486
+
487
+ def _summarize(ap=1, iouThr=None, areaRng="all", maxDets=100):
488
+ p = self.params
489
+ iStr = " {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}"
490
+ titleStr = "Average Precision" if ap == 1 else "Average Recall"
491
+ typeStr = "(AP)" if ap == 1 else "(AR)"
492
+ iouStr = (
493
+ "{:0.2f}:{:0.2f}".format(p.iouThrs[0], p.iouThrs[-1])
494
+ if iouThr is None
495
+ else "{:0.2f}".format(iouThr)
496
+ )
497
+
498
+ aind = [i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng]
499
+ mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets]
500
+ if ap == 1:
501
+ # dimension of precision: [TxRxKxAxM]
502
+ s = self.eval["precision"]
503
+ # IoU
504
+ if iouThr is not None:
505
+ t = np.where(iouThr == p.iouThrs)[0]
506
+ s = s[t]
507
+ s = s[:, :, :, aind, mind]
508
+ else:
509
+ # dimension of recall: [TxKxAxM]
510
+ s = self.eval["recall"]
511
+ if iouThr is not None:
512
+ t = np.where(iouThr == p.iouThrs)[0]
513
+ s = s[t]
514
+ s = s[:, :, aind, mind]
515
+ if len(s[s > -1]) == 0:
516
+ mean_s = -1
517
+ else:
518
+ mean_s = np.mean(s[s > -1])
519
+ print(iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets, mean_s))
520
+ return mean_s
521
+
522
+ def _summarizeDets():
523
+ stats = np.zeros((12,))
524
+ # Evaluate AP using the custom limit on maximum detections per image
525
+ stats[0] = _summarize(1, maxDets=self.params.maxDets[2])
526
+ stats[1] = _summarize(1, iouThr=0.5, maxDets=self.params.maxDets[2])
527
+ stats[2] = _summarize(1, iouThr=0.75, maxDets=self.params.maxDets[2])
528
+ stats[3] = _summarize(1, areaRng="small", maxDets=self.params.maxDets[2])
529
+ stats[4] = _summarize(1, areaRng="medium", maxDets=self.params.maxDets[2])
530
+ stats[5] = _summarize(1, areaRng="large", maxDets=self.params.maxDets[2])
531
+ stats[6] = _summarize(0, maxDets=self.params.maxDets[0])
532
+ stats[7] = _summarize(0, maxDets=self.params.maxDets[1])
533
+ stats[8] = _summarize(0, maxDets=self.params.maxDets[2])
534
+ stats[9] = _summarize(0, areaRng="small", maxDets=self.params.maxDets[2])
535
+ stats[10] = _summarize(0, areaRng="medium", maxDets=self.params.maxDets[2])
536
+ stats[11] = _summarize(0, areaRng="large", maxDets=self.params.maxDets[2])
537
+ return stats
538
+
539
+ def _summarizeKps():
540
+ stats = np.zeros((10,))
541
+ stats[0] = _summarize(1, maxDets=20)
542
+ stats[1] = _summarize(1, maxDets=20, iouThr=0.5)
543
+ stats[2] = _summarize(1, maxDets=20, iouThr=0.75)
544
+ stats[3] = _summarize(1, maxDets=20, areaRng="medium")
545
+ stats[4] = _summarize(1, maxDets=20, areaRng="large")
546
+ stats[5] = _summarize(0, maxDets=20)
547
+ stats[6] = _summarize(0, maxDets=20, iouThr=0.5)
548
+ stats[7] = _summarize(0, maxDets=20, iouThr=0.75)
549
+ stats[8] = _summarize(0, maxDets=20, areaRng="medium")
550
+ stats[9] = _summarize(0, maxDets=20, areaRng="large")
551
+ return stats
552
+
553
+ if not self.eval:
554
+ raise Exception("Please run accumulate() first")
555
+ iouType = self.params.iouType
556
+ if iouType == "segm":
557
+ summarize = _summarizeDets
558
+ elif iouType == "keypoints":
559
+ summarize = _summarizeKps
560
+ self.stats = summarize()
561
+
562
+ def __str__(self):
563
+ self.summarize()
oneformer/evaluation/detection_coco_evaluator.py ADDED
@@ -0,0 +1,723 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/coco_evaluation.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import contextlib
7
+ import copy
8
+ import io
9
+ import itertools
10
+ import json
11
+ import logging
12
+ import numpy as np
13
+ import os
14
+ import pickle
15
+ from collections import OrderedDict
16
+ import pycocotools.mask as mask_util
17
+ import torch
18
+ from pycocotools.coco import COCO
19
+ from pycocotools.cocoeval import COCOeval
20
+ from tabulate import tabulate
21
+
22
+ import detectron2.utils.comm as comm
23
+ from detectron2.config import CfgNode
24
+ from detectron2.data import MetadataCatalog
25
+ from detectron2.data.datasets.coco import convert_to_coco_json
26
+ from detectron2.structures import Boxes, BoxMode, pairwise_iou
27
+ from detectron2.utils.file_io import PathManager
28
+ from detectron2.utils.logger import create_small_table
29
+
30
+ from .evaluator import DatasetEvaluator
31
+
32
+ try:
33
+ from detectron2.evaluation.fast_eval_api import COCOeval_opt
34
+ except ImportError:
35
+ COCOeval_opt = COCOeval
36
+
37
+
38
+ class DetectionCOCOEvaluator(DatasetEvaluator):
39
+ """
40
+ Evaluate AR for object proposals, AP for instance detection/segmentation, AP
41
+ for keypoint detection outputs using COCO's metrics.
42
+ See http://cocodataset.org/#detection-eval and
43
+ http://cocodataset.org/#keypoints-eval to understand its metrics.
44
+ The metrics range from 0 to 100 (instead of 0 to 1), where a -1 or NaN means
45
+ the metric cannot be computed (e.g. due to no predictions made).
46
+
47
+ In addition to COCO, this evaluator is able to support any bounding box detection,
48
+ instance segmentation, or keypoint detection dataset.
49
+ """
50
+
51
+ def __init__(
52
+ self,
53
+ dataset_name,
54
+ tasks=None,
55
+ distributed=True,
56
+ output_dir=None,
57
+ *,
58
+ max_dets_per_image=None,
59
+ use_fast_impl=True,
60
+ kpt_oks_sigmas=(),
61
+ allow_cached_coco=True,
62
+ ):
63
+ """
64
+ Args:
65
+ dataset_name (str): name of the dataset to be evaluated.
66
+ It must have either the following corresponding metadata:
67
+
68
+ "json_file": the path to the COCO format annotation
69
+
70
+ Or it must be in detectron2's standard dataset format
71
+ so it can be converted to COCO format automatically.
72
+ tasks (tuple[str]): tasks that can be evaluated under the given
73
+ configuration. A task is one of "bbox", "segm", "keypoints".
74
+ By default, will infer this automatically from predictions.
75
+ distributed (True): if True, will collect results from all ranks and run evaluation
76
+ in the main process.
77
+ Otherwise, will only evaluate the results in the current process.
78
+ output_dir (str): optional, an output directory to dump all
79
+ results predicted on the dataset. The dump contains two files:
80
+
81
+ 1. "instances_predictions.pth" a file that can be loaded with `torch.load` and
82
+ contains all the results in the format they are produced by the model.
83
+ 2. "coco_instances_results.json" a json file in COCO's result format.
84
+ max_dets_per_image (int): limit on the maximum number of detections per image.
85
+ By default in COCO, this limit is to 100, but this can be customized
86
+ to be greater, as is needed in evaluation metrics AP fixed and AP pool
87
+ (see https://arxiv.org/pdf/2102.01066.pdf)
88
+ This doesn't affect keypoint evaluation.
89
+ use_fast_impl (bool): use a fast but **unofficial** implementation to compute AP.
90
+ Although the results should be very close to the official implementation in COCO
91
+ API, it is still recommended to compute results with the official API for use in
92
+ papers. The faster implementation also uses more RAM.
93
+ kpt_oks_sigmas (list[float]): The sigmas used to calculate keypoint OKS.
94
+ See http://cocodataset.org/#keypoints-eval
95
+ When empty, it will use the defaults in COCO.
96
+ Otherwise it should be the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS.
97
+ allow_cached_coco (bool): Whether to use cached coco json from previous validation
98
+ runs. You should set this to False if you need to use different validation data.
99
+ Defaults to True.
100
+ """
101
+ self._logger = logging.getLogger(__name__)
102
+ self._distributed = distributed
103
+ self._output_dir = output_dir
104
+
105
+ if use_fast_impl and (COCOeval_opt is COCOeval):
106
+ self._logger.info("Fast COCO eval is not built. Falling back to official COCO eval.")
107
+ use_fast_impl = False
108
+ self._use_fast_impl = use_fast_impl
109
+
110
+ # COCOeval requires the limit on the number of detections per image (maxDets) to be a list
111
+ # with at least 3 elements. The default maxDets in COCOeval is [1, 10, 100], in which the
112
+ # 3rd element (100) is used as the limit on the number of detections per image when
113
+ # evaluating AP. COCOEvaluator expects an integer for max_dets_per_image, so for COCOeval,
114
+ # we reformat max_dets_per_image into [1, 10, max_dets_per_image], based on the defaults.
115
+ if max_dets_per_image is None:
116
+ max_dets_per_image = [1, 10, 100]
117
+ else:
118
+ max_dets_per_image = [1, 10, max_dets_per_image]
119
+ self._max_dets_per_image = max_dets_per_image
120
+
121
+ if tasks is not None and isinstance(tasks, CfgNode):
122
+ kpt_oks_sigmas = (
123
+ tasks.TEST.KEYPOINT_OKS_SIGMAS if not kpt_oks_sigmas else kpt_oks_sigmas
124
+ )
125
+ self._logger.warn(
126
+ "COCO Evaluator instantiated using config, this is deprecated behavior."
127
+ " Please pass in explicit arguments instead."
128
+ )
129
+ self._tasks = None # Infering it from predictions should be better
130
+ else:
131
+ self._tasks = tasks
132
+
133
+ self._cpu_device = torch.device("cpu")
134
+
135
+ self._metadata = MetadataCatalog.get(dataset_name)
136
+ if not hasattr(self._metadata, "json_file"):
137
+ if output_dir is None:
138
+ raise ValueError(
139
+ "output_dir must be provided to COCOEvaluator "
140
+ "for datasets not in COCO format."
141
+ )
142
+ self._logger.info(f"Trying to convert '{dataset_name}' to COCO format ...")
143
+
144
+ cache_path = os.path.join(output_dir, f"{dataset_name}_coco_format.json")
145
+ self._metadata.json_file = cache_path
146
+ convert_to_coco_json(dataset_name, cache_path, allow_cached=allow_cached_coco)
147
+
148
+ json_file = PathManager.get_local_path(self._metadata.json_file)
149
+ with contextlib.redirect_stdout(io.StringIO()):
150
+ self._coco_api = COCO(json_file)
151
+
152
+ # Test set json files do not contain annotations (evaluation must be
153
+ # performed using the COCO evaluation server).
154
+ self._do_evaluation = "annotations" in self._coco_api.dataset
155
+ if self._do_evaluation:
156
+ self._kpt_oks_sigmas = kpt_oks_sigmas
157
+
158
+ def reset(self):
159
+ self._predictions = []
160
+
161
+ def process(self, inputs, outputs):
162
+ """
163
+ Args:
164
+ inputs: the inputs to a COCO model (e.g., GeneralizedRCNN).
165
+ It is a list of dict. Each dict corresponds to an image and
166
+ contains keys like "height", "width", "file_name", "image_id".
167
+ outputs: the outputs of a COCO model. It is a list of dicts with key
168
+ "box_instances" that contains :class:`Instances`.
169
+ """
170
+ for input, output in zip(inputs, outputs):
171
+ prediction = {"image_id": input["image_id"]}
172
+
173
+ if "box_instances" in output:
174
+ instances = output["box_instances"].to(self._cpu_device)
175
+ prediction["box_instances"] = instances_to_coco_json(instances, input["image_id"])
176
+ if "proposals" in output:
177
+ prediction["proposals"] = output["proposals"].to(self._cpu_device)
178
+ if len(prediction) > 1:
179
+ self._predictions.append(prediction)
180
+
181
+ def evaluate(self, img_ids=None):
182
+ """
183
+ Args:
184
+ img_ids: a list of image IDs to evaluate on. Default to None for the whole dataset
185
+ """
186
+ if self._distributed:
187
+ comm.synchronize()
188
+ predictions = comm.gather(self._predictions, dst=0)
189
+ predictions = list(itertools.chain(*predictions))
190
+
191
+ if not comm.is_main_process():
192
+ return {}
193
+ else:
194
+ predictions = self._predictions
195
+
196
+ if len(predictions) == 0:
197
+ self._logger.warning("[COCOEvaluator] Did not receive valid predictions.")
198
+ return {}
199
+
200
+ if self._output_dir:
201
+ PathManager.mkdirs(self._output_dir)
202
+ file_path = os.path.join(self._output_dir, "instances_predictions.pth")
203
+ with PathManager.open(file_path, "wb") as f:
204
+ torch.save(predictions, f)
205
+
206
+ self._results = OrderedDict()
207
+ if "proposals" in predictions[0]:
208
+ self._eval_box_proposals(predictions)
209
+ if "box_instances" in predictions[0]:
210
+ self._eval_predictions(predictions, img_ids=img_ids)
211
+ # Copy so the caller can do whatever with results
212
+ return copy.deepcopy(self._results)
213
+
214
+ def _tasks_from_predictions(self, predictions):
215
+ """
216
+ Get COCO API "tasks" (i.e. iou_type) from COCO-format predictions.
217
+ """
218
+ tasks = {"bbox"}
219
+ for pred in predictions:
220
+ if "keypoints" in pred:
221
+ tasks.add("keypoints")
222
+ return sorted(tasks)
223
+
224
+ def _eval_predictions(self, predictions, img_ids=None):
225
+ """
226
+ Evaluate predictions. Fill self._results with the metrics of the tasks.
227
+ """
228
+ self._logger.info("Preparing results for COCO format ...")
229
+ coco_results = list(itertools.chain(*[x["box_instances"] for x in predictions]))
230
+ tasks = self._tasks or self._tasks_from_predictions(coco_results)
231
+
232
+ # unmap the category ids for COCO
233
+ if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
234
+ dataset_id_to_contiguous_id = self._metadata.thing_dataset_id_to_contiguous_id
235
+ all_contiguous_ids = list(dataset_id_to_contiguous_id.values())
236
+ num_classes = len(all_contiguous_ids)
237
+ assert min(all_contiguous_ids) == 0 and max(all_contiguous_ids) == num_classes - 1
238
+
239
+ reverse_id_mapping = {v: k for k, v in dataset_id_to_contiguous_id.items()}
240
+ for result in coco_results:
241
+ category_id = result["category_id"]
242
+ assert category_id < num_classes, (
243
+ f"A prediction has class={category_id}, "
244
+ f"but the dataset only has {num_classes} classes and "
245
+ f"predicted class id should be in [0, {num_classes - 1}]."
246
+ )
247
+ result["category_id"] = reverse_id_mapping[category_id]
248
+
249
+ if self._output_dir:
250
+ file_path = os.path.join(self._output_dir, "coco_instances_results.json")
251
+ self._logger.info("Saving results to {}".format(file_path))
252
+ with PathManager.open(file_path, "w") as f:
253
+ f.write(json.dumps(coco_results))
254
+ f.flush()
255
+
256
+ if not self._do_evaluation:
257
+ self._logger.info("Annotations are not available for evaluation.")
258
+ return
259
+
260
+ self._logger.info(
261
+ "Evaluating predictions with {} COCO API...".format(
262
+ "unofficial" if self._use_fast_impl else "official"
263
+ )
264
+ )
265
+ for task in sorted(tasks):
266
+ assert task in {"bbox", "keypoints"}, f"Got unknown task: {task}!"
267
+ coco_eval = (
268
+ _evaluate_predictions_on_coco(
269
+ self._coco_api,
270
+ coco_results,
271
+ task,
272
+ kpt_oks_sigmas=self._kpt_oks_sigmas,
273
+ use_fast_impl=self._use_fast_impl,
274
+ img_ids=img_ids,
275
+ max_dets_per_image=self._max_dets_per_image,
276
+ )
277
+ if len(coco_results) > 0
278
+ else None # cocoapi does not handle empty results very well
279
+ )
280
+
281
+ res = self._derive_coco_results(
282
+ coco_eval, task, class_names=self._metadata.get("thing_classes")
283
+ )
284
+ self._results[task] = res
285
+
286
+ def _eval_box_proposals(self, predictions):
287
+ """
288
+ Evaluate the box proposals in predictions.
289
+ Fill self._results with the metrics for "box_proposals" task.
290
+ """
291
+ if self._output_dir:
292
+ # Saving generated box proposals to file.
293
+ # Predicted box_proposals are in XYXY_ABS mode.
294
+ bbox_mode = BoxMode.XYXY_ABS.value
295
+ ids, boxes, objectness_logits = [], [], []
296
+ for prediction in predictions:
297
+ ids.append(prediction["image_id"])
298
+ boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy())
299
+ objectness_logits.append(prediction["proposals"].objectness_logits.numpy())
300
+
301
+ proposal_data = {
302
+ "boxes": boxes,
303
+ "objectness_logits": objectness_logits,
304
+ "ids": ids,
305
+ "bbox_mode": bbox_mode,
306
+ }
307
+ with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f:
308
+ pickle.dump(proposal_data, f)
309
+
310
+ if not self._do_evaluation:
311
+ self._logger.info("Annotations are not available for evaluation.")
312
+ return
313
+
314
+ self._logger.info("Evaluating bbox proposals ...")
315
+ res = {}
316
+ areas = {"all": "", "small": "s", "medium": "m", "large": "l"}
317
+ for limit in [100, 1000]:
318
+ for area, suffix in areas.items():
319
+ stats = _evaluate_box_proposals(predictions, self._coco_api, area=area, limit=limit)
320
+ key = "AR{}@{:d}".format(suffix, limit)
321
+ res[key] = float(stats["ar"].item() * 100)
322
+ self._logger.info("Proposal metrics: \n" + create_small_table(res))
323
+ self._results["box_proposals"] = res
324
+
325
+ def _derive_coco_results(self, coco_eval, iou_type, class_names=None):
326
+ """
327
+ Derive the desired score numbers from summarized COCOeval.
328
+
329
+ Args:
330
+ coco_eval (None or COCOEval): None represents no predictions from model.
331
+ iou_type (str):
332
+ class_names (None or list[str]): if provided, will use it to predict
333
+ per-category AP.
334
+
335
+ Returns:
336
+ a dict of {metric name: score}
337
+ """
338
+
339
+ metrics = {
340
+ "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl"],
341
+ "keypoints": ["AP", "AP50", "AP75", "APm", "APl"],
342
+ }[iou_type]
343
+
344
+ if coco_eval is None:
345
+ self._logger.warn("No predictions from the model!")
346
+ return {metric: float("nan") for metric in metrics}
347
+
348
+ # the standard metrics
349
+ results = {
350
+ metric: float(coco_eval.stats[idx] * 100 if coco_eval.stats[idx] >= 0 else "nan")
351
+ for idx, metric in enumerate(metrics)
352
+ }
353
+ self._logger.info(
354
+ "Evaluation results for {}: \n".format(iou_type) + create_small_table(results)
355
+ )
356
+ if not np.isfinite(sum(results.values())):
357
+ self._logger.info("Some metrics cannot be computed and is shown as NaN.")
358
+
359
+ if class_names is None or len(class_names) <= 1:
360
+ return results
361
+ # Compute per-category AP
362
+ # from https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L222-L252 # noqa
363
+ precisions = coco_eval.eval["precision"]
364
+ # precision has dims (iou, recall, cls, area range, max dets)
365
+ assert len(class_names) == precisions.shape[2]
366
+
367
+ results_per_category = []
368
+ for idx, name in enumerate(class_names):
369
+ # area range index 0: all area ranges
370
+ # max dets index -1: typically 100 per image
371
+ precision = precisions[:, :, idx, 0, -1]
372
+ precision = precision[precision > -1]
373
+ ap = np.mean(precision) if precision.size else float("nan")
374
+ results_per_category.append(("{}".format(name), float(ap * 100)))
375
+
376
+ # tabulate it
377
+ N_COLS = min(6, len(results_per_category) * 2)
378
+ results_flatten = list(itertools.chain(*results_per_category))
379
+ results_2d = itertools.zip_longest(*[results_flatten[i::N_COLS] for i in range(N_COLS)])
380
+ table = tabulate(
381
+ results_2d,
382
+ tablefmt="pipe",
383
+ floatfmt=".3f",
384
+ headers=["category", "AP"] * (N_COLS // 2),
385
+ numalign="left",
386
+ )
387
+ self._logger.info("Per-category {} AP: \n".format(iou_type) + table)
388
+
389
+ results.update({"AP-" + name: ap for name, ap in results_per_category})
390
+ return results
391
+
392
+
393
+ def instances_to_coco_json(instances, img_id):
394
+ """
395
+ Dump an "Instances" object to a COCO-format json that's used for evaluation.
396
+
397
+ Args:
398
+ instances (Instances):
399
+ img_id (int): the image id
400
+
401
+ Returns:
402
+ list[dict]: list of json annotations in COCO format.
403
+ """
404
+ num_instance = len(instances)
405
+ if num_instance == 0:
406
+ return []
407
+
408
+ boxes = instances.pred_boxes.tensor.numpy()
409
+ boxes = BoxMode.convert(boxes, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
410
+ boxes = boxes.tolist()
411
+ scores = instances.scores.tolist()
412
+ classes = instances.pred_classes.tolist()
413
+
414
+ has_mask = instances.has("pred_masks")
415
+ if has_mask:
416
+ # use RLE to encode the masks, because they are too large and takes memory
417
+ # since this evaluator stores outputs of the entire dataset
418
+ rles = [
419
+ mask_util.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0]
420
+ for mask in instances.pred_masks
421
+ ]
422
+ for rle in rles:
423
+ # "counts" is an array encoded by mask_util as a byte-stream. Python3's
424
+ # json writer which always produces strings cannot serialize a bytestream
425
+ # unless you decode it. Thankfully, utf-8 works out (which is also what
426
+ # the pycocotools/_mask.pyx does).
427
+ rle["counts"] = rle["counts"].decode("utf-8")
428
+
429
+ has_keypoints = instances.has("pred_keypoints")
430
+ if has_keypoints:
431
+ keypoints = instances.pred_keypoints
432
+
433
+ results = []
434
+ for k in range(num_instance):
435
+ result = {
436
+ "image_id": img_id,
437
+ "category_id": classes[k],
438
+ "bbox": boxes[k],
439
+ "score": scores[k],
440
+ }
441
+ if has_mask:
442
+ result["segmentation"] = rles[k]
443
+ if has_keypoints:
444
+ # In COCO annotations,
445
+ # keypoints coordinates are pixel indices.
446
+ # However our predictions are floating point coordinates.
447
+ # Therefore we subtract 0.5 to be consistent with the annotation format.
448
+ # This is the inverse of data loading logic in `datasets/coco.py`.
449
+ keypoints[k][:, :2] -= 0.5
450
+ result["keypoints"] = keypoints[k].flatten().tolist()
451
+ results.append(result)
452
+ return results
453
+
454
+
455
+ # inspired from Detectron:
456
+ # https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa
457
+ def _evaluate_box_proposals(dataset_predictions, coco_api, thresholds=None, area="all", limit=None):
458
+ """
459
+ Evaluate detection proposal recall metrics. This function is a much
460
+ faster alternative to the official COCO API recall evaluation code. However,
461
+ it produces slightly different results.
462
+ """
463
+ # Record max overlap value for each gt box
464
+ # Return vector of overlap values
465
+ areas = {
466
+ "all": 0,
467
+ "small": 1,
468
+ "medium": 2,
469
+ "large": 3,
470
+ "96-128": 4,
471
+ "128-256": 5,
472
+ "256-512": 6,
473
+ "512-inf": 7,
474
+ }
475
+ area_ranges = [
476
+ [0**2, 1e5**2], # all
477
+ [0**2, 32**2], # small
478
+ [32**2, 96**2], # medium
479
+ [96**2, 1e5**2], # large
480
+ [96**2, 128**2], # 96-128
481
+ [128**2, 256**2], # 128-256
482
+ [256**2, 512**2], # 256-512
483
+ [512**2, 1e5**2],
484
+ ] # 512-inf
485
+ assert area in areas, "Unknown area range: {}".format(area)
486
+ area_range = area_ranges[areas[area]]
487
+ gt_overlaps = []
488
+ num_pos = 0
489
+
490
+ for prediction_dict in dataset_predictions:
491
+ predictions = prediction_dict["proposals"]
492
+
493
+ # sort predictions in descending order
494
+ # TODO maybe remove this and make it explicit in the documentation
495
+ inds = predictions.objectness_logits.sort(descending=True)[1]
496
+ predictions = predictions[inds]
497
+
498
+ ann_ids = coco_api.getAnnIds(imgIds=prediction_dict["image_id"])
499
+ anno = coco_api.loadAnns(ann_ids)
500
+ gt_boxes = [
501
+ BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS)
502
+ for obj in anno
503
+ if obj["iscrowd"] == 0
504
+ ]
505
+ gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4) # guard against no boxes
506
+ gt_boxes = Boxes(gt_boxes)
507
+ gt_areas = torch.as_tensor([obj["area"] for obj in anno if obj["iscrowd"] == 0])
508
+
509
+ if len(gt_boxes) == 0 or len(predictions) == 0:
510
+ continue
511
+
512
+ valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1])
513
+ gt_boxes = gt_boxes[valid_gt_inds]
514
+
515
+ num_pos += len(gt_boxes)
516
+
517
+ if len(gt_boxes) == 0:
518
+ continue
519
+
520
+ if limit is not None and len(predictions) > limit:
521
+ predictions = predictions[:limit]
522
+
523
+ overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes)
524
+
525
+ _gt_overlaps = torch.zeros(len(gt_boxes))
526
+ for j in range(min(len(predictions), len(gt_boxes))):
527
+ # find which proposal box maximally covers each gt box
528
+ # and get the iou amount of coverage for each gt box
529
+ max_overlaps, argmax_overlaps = overlaps.max(dim=0)
530
+
531
+ # find which gt box is 'best' covered (i.e. 'best' = most iou)
532
+ gt_ovr, gt_ind = max_overlaps.max(dim=0)
533
+ assert gt_ovr >= 0
534
+ # find the proposal box that covers the best covered gt box
535
+ box_ind = argmax_overlaps[gt_ind]
536
+ # record the iou coverage of this gt box
537
+ _gt_overlaps[j] = overlaps[box_ind, gt_ind]
538
+ assert _gt_overlaps[j] == gt_ovr
539
+ # mark the proposal box and the gt box as used
540
+ overlaps[box_ind, :] = -1
541
+ overlaps[:, gt_ind] = -1
542
+
543
+ # append recorded iou coverage level
544
+ gt_overlaps.append(_gt_overlaps)
545
+ gt_overlaps = (
546
+ torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32)
547
+ )
548
+ gt_overlaps, _ = torch.sort(gt_overlaps)
549
+
550
+ if thresholds is None:
551
+ step = 0.05
552
+ thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
553
+ recalls = torch.zeros_like(thresholds)
554
+ # compute recall for each iou threshold
555
+ for i, t in enumerate(thresholds):
556
+ recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
557
+ # ar = 2 * np.trapz(recalls, thresholds)
558
+ ar = recalls.mean()
559
+ return {
560
+ "ar": ar,
561
+ "recalls": recalls,
562
+ "thresholds": thresholds,
563
+ "gt_overlaps": gt_overlaps,
564
+ "num_pos": num_pos,
565
+ }
566
+
567
+
568
+ def _evaluate_predictions_on_coco(
569
+ coco_gt,
570
+ coco_results,
571
+ iou_type,
572
+ kpt_oks_sigmas=None,
573
+ use_fast_impl=True,
574
+ img_ids=None,
575
+ max_dets_per_image=None,
576
+ ):
577
+ """
578
+ Evaluate the coco results using COCOEval API.
579
+ """
580
+ assert len(coco_results) > 0
581
+
582
+ if iou_type == "segm":
583
+ coco_results = copy.deepcopy(coco_results)
584
+ # When evaluating mask AP, if the results contain bbox, cocoapi will
585
+ # use the box area as the area of the instance, instead of the mask area.
586
+ # This leads to a different definition of small/medium/large.
587
+ # We remove the bbox field to let mask AP use mask area.
588
+ for c in coco_results:
589
+ c.pop("bbox", None)
590
+
591
+ coco_dt = coco_gt.loadRes(coco_results)
592
+ coco_eval = (COCOeval_opt if use_fast_impl else COCOeval)(coco_gt, coco_dt, iou_type)
593
+ # For COCO, the default max_dets_per_image is [1, 10, 100].
594
+ if max_dets_per_image is None:
595
+ max_dets_per_image = [1, 10, 100] # Default from COCOEval
596
+ else:
597
+ assert (
598
+ len(max_dets_per_image) >= 3
599
+ ), "COCOeval requires maxDets (and max_dets_per_image) to have length at least 3"
600
+ # In the case that user supplies a custom input for max_dets_per_image,
601
+ # apply COCOevalMaxDets to evaluate AP with the custom input.
602
+ if max_dets_per_image[2] != 100:
603
+ coco_eval = COCOevalMaxDets(coco_gt, coco_dt, iou_type)
604
+ if iou_type != "keypoints":
605
+ coco_eval.params.maxDets = max_dets_per_image
606
+
607
+ if img_ids is not None:
608
+ coco_eval.params.imgIds = img_ids
609
+
610
+ if iou_type == "keypoints":
611
+ # Use the COCO default keypoint OKS sigmas unless overrides are specified
612
+ if kpt_oks_sigmas:
613
+ assert hasattr(coco_eval.params, "kpt_oks_sigmas"), "pycocotools is too old!"
614
+ coco_eval.params.kpt_oks_sigmas = np.array(kpt_oks_sigmas)
615
+ # COCOAPI requires every detection and every gt to have keypoints, so
616
+ # we just take the first entry from both
617
+ num_keypoints_dt = len(coco_results[0]["keypoints"]) // 3
618
+ num_keypoints_gt = len(next(iter(coco_gt.anns.values()))["keypoints"]) // 3
619
+ num_keypoints_oks = len(coco_eval.params.kpt_oks_sigmas)
620
+ assert num_keypoints_oks == num_keypoints_dt == num_keypoints_gt, (
621
+ f"[COCOEvaluator] Prediction contain {num_keypoints_dt} keypoints. "
622
+ f"Ground truth contains {num_keypoints_gt} keypoints. "
623
+ f"The length of cfg.TEST.KEYPOINT_OKS_SIGMAS is {num_keypoints_oks}. "
624
+ "They have to agree with each other. For meaning of OKS, please refer to "
625
+ "http://cocodataset.org/#keypoints-eval."
626
+ )
627
+
628
+ coco_eval.evaluate()
629
+ coco_eval.accumulate()
630
+ coco_eval.summarize()
631
+
632
+ return coco_eval
633
+
634
+
635
+ class COCOevalMaxDets(COCOeval):
636
+ """
637
+ Modified version of COCOeval for evaluating AP with a custom
638
+ maxDets (by default for COCO, maxDets is 100)
639
+ """
640
+
641
+ def summarize(self):
642
+ """
643
+ Compute and display summary metrics for evaluation results given
644
+ a custom value for max_dets_per_image
645
+ """
646
+
647
+ def _summarize(ap=1, iouThr=None, areaRng="all", maxDets=100):
648
+ p = self.params
649
+ iStr = " {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}"
650
+ titleStr = "Average Precision" if ap == 1 else "Average Recall"
651
+ typeStr = "(AP)" if ap == 1 else "(AR)"
652
+ iouStr = (
653
+ "{:0.2f}:{:0.2f}".format(p.iouThrs[0], p.iouThrs[-1])
654
+ if iouThr is None
655
+ else "{:0.2f}".format(iouThr)
656
+ )
657
+
658
+ aind = [i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng]
659
+ mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets]
660
+ if ap == 1:
661
+ # dimension of precision: [TxRxKxAxM]
662
+ s = self.eval["precision"]
663
+ # IoU
664
+ if iouThr is not None:
665
+ t = np.where(iouThr == p.iouThrs)[0]
666
+ s = s[t]
667
+ s = s[:, :, :, aind, mind]
668
+ else:
669
+ # dimension of recall: [TxKxAxM]
670
+ s = self.eval["recall"]
671
+ if iouThr is not None:
672
+ t = np.where(iouThr == p.iouThrs)[0]
673
+ s = s[t]
674
+ s = s[:, :, aind, mind]
675
+ if len(s[s > -1]) == 0:
676
+ mean_s = -1
677
+ else:
678
+ mean_s = np.mean(s[s > -1])
679
+ print(iStr.format(titleStr, typeStr, iouStr, areaRng, maxDets, mean_s))
680
+ return mean_s
681
+
682
+ def _summarizeDets():
683
+ stats = np.zeros((12,))
684
+ # Evaluate AP using the custom limit on maximum detections per image
685
+ stats[0] = _summarize(1, maxDets=self.params.maxDets[2])
686
+ stats[1] = _summarize(1, iouThr=0.5, maxDets=self.params.maxDets[2])
687
+ stats[2] = _summarize(1, iouThr=0.75, maxDets=self.params.maxDets[2])
688
+ stats[3] = _summarize(1, areaRng="small", maxDets=self.params.maxDets[2])
689
+ stats[4] = _summarize(1, areaRng="medium", maxDets=self.params.maxDets[2])
690
+ stats[5] = _summarize(1, areaRng="large", maxDets=self.params.maxDets[2])
691
+ stats[6] = _summarize(0, maxDets=self.params.maxDets[0])
692
+ stats[7] = _summarize(0, maxDets=self.params.maxDets[1])
693
+ stats[8] = _summarize(0, maxDets=self.params.maxDets[2])
694
+ stats[9] = _summarize(0, areaRng="small", maxDets=self.params.maxDets[2])
695
+ stats[10] = _summarize(0, areaRng="medium", maxDets=self.params.maxDets[2])
696
+ stats[11] = _summarize(0, areaRng="large", maxDets=self.params.maxDets[2])
697
+ return stats
698
+
699
+ def _summarizeKps():
700
+ stats = np.zeros((10,))
701
+ stats[0] = _summarize(1, maxDets=20)
702
+ stats[1] = _summarize(1, maxDets=20, iouThr=0.5)
703
+ stats[2] = _summarize(1, maxDets=20, iouThr=0.75)
704
+ stats[3] = _summarize(1, maxDets=20, areaRng="medium")
705
+ stats[4] = _summarize(1, maxDets=20, areaRng="large")
706
+ stats[5] = _summarize(0, maxDets=20)
707
+ stats[6] = _summarize(0, maxDets=20, iouThr=0.5)
708
+ stats[7] = _summarize(0, maxDets=20, iouThr=0.75)
709
+ stats[8] = _summarize(0, maxDets=20, areaRng="medium")
710
+ stats[9] = _summarize(0, maxDets=20, areaRng="large")
711
+ return stats
712
+
713
+ if not self.eval:
714
+ raise Exception("Please run accumulate() first")
715
+ iouType = self.params.iouType
716
+ if iouType == "segm" or iouType == "bbox":
717
+ summarize = _summarizeDets
718
+ elif iouType == "keypoints":
719
+ summarize = _summarizeKps
720
+ self.stats = summarize()
721
+
722
+ def __str__(self):
723
+ self.summarize()
oneformer/evaluation/evaluator.py ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/evaluator.py
3
+ # Modified by Jitesh Jain (https://github.com/praeclarumjj3)
4
+ # ------------------------------------------------------------------------------
5
+
6
+ import datetime
7
+ import logging
8
+ import time
9
+ from collections import OrderedDict, abc
10
+ from contextlib import ExitStack, contextmanager
11
+ from typing import List, Union
12
+ import torch
13
+ from torch import nn
14
+
15
+ from detectron2.utils.comm import get_world_size, is_main_process
16
+ from detectron2.utils.logger import log_every_n_seconds
17
+
18
+
19
+ class DatasetEvaluator:
20
+ """
21
+ Base class for a dataset evaluator.
22
+
23
+ The function :func:`inference_on_dataset` runs the model over
24
+ all samples in the dataset, and have a DatasetEvaluator to process the inputs/outputs.
25
+
26
+ This class will accumulate information of the inputs/outputs (by :meth:`process`),
27
+ and produce evaluation results in the end (by :meth:`evaluate`).
28
+ """
29
+
30
+ def reset(self):
31
+ """
32
+ Preparation for a new round of evaluation.
33
+ Should be called before starting a round of evaluation.
34
+ """
35
+ pass
36
+
37
+ def process(self, inputs, outputs):
38
+ """
39
+ Process the pair of inputs and outputs.
40
+ If they contain batches, the pairs can be consumed one-by-one using `zip`:
41
+
42
+ .. code-block:: python
43
+
44
+ for input_, output in zip(inputs, outputs):
45
+ # do evaluation on single input/output pair
46
+ ...
47
+
48
+ Args:
49
+ inputs (list): the inputs that's used to call the model.
50
+ outputs (list): the return value of `model(inputs)`
51
+ """
52
+ pass
53
+
54
+ def evaluate(self):
55
+ """
56
+ Evaluate/summarize the performance, after processing all input/output pairs.
57
+
58
+ Returns:
59
+ dict:
60
+ A new evaluator class can return a dict of arbitrary format
61
+ as long as the user can process the results.
62
+ In our train_net.py, we expect the following format:
63
+
64
+ * key: the name of the task (e.g., bbox)
65
+ * value: a dict of {metric name: score}, e.g.: {"AP50": 80}
66
+ """
67
+ pass
68
+
69
+
70
+ class DatasetEvaluators(DatasetEvaluator):
71
+ """
72
+ Wrapper class to combine multiple :class:`DatasetEvaluator` instances.
73
+
74
+ This class dispatches every evaluation call to
75
+ all of its :class:`DatasetEvaluator`.
76
+ """
77
+
78
+ def __init__(self, evaluators):
79
+ """
80
+ Args:
81
+ evaluators (list): the evaluators to combine.
82
+ """
83
+ super().__init__()
84
+ self._evaluators = evaluators
85
+
86
+ def reset(self):
87
+ for evaluator in self._evaluators:
88
+ evaluator.reset()
89
+
90
+ def process(self, inputs, outputs):
91
+ for evaluator in self._evaluators:
92
+ evaluator.process(inputs, outputs)
93
+
94
+ def evaluate(self):
95
+ results = OrderedDict()
96
+ for evaluator in self._evaluators:
97
+ result = evaluator.evaluate()
98
+ if is_main_process() and result is not None:
99
+ for k, v in result.items():
100
+ assert (
101
+ k not in results
102
+ ), "Different evaluators produce results with the same key {}".format(k)
103
+ results[k] = v
104
+ return results
105
+
106
+
107
+ def inference_on_dataset(
108
+ model, data_loader, evaluator: Union[DatasetEvaluator, List[DatasetEvaluator], None]
109
+ ):
110
+ """
111
+ Run model on the data_loader and evaluate the metrics with evaluator.
112
+ Also benchmark the inference speed of `model.__call__` accurately.
113
+ The model will be used in eval mode.
114
+
115
+ Args:
116
+ model (callable): a callable which takes an object from
117
+ `data_loader` and returns some outputs.
118
+
119
+ If it's an nn.Module, it will be temporarily set to `eval` mode.
120
+ If you wish to evaluate a model in `training` mode instead, you can
121
+ wrap the given model and override its behavior of `.eval()` and `.train()`.
122
+ data_loader: an iterable object with a length.
123
+ The elements it generates will be the inputs to the model.
124
+ evaluator: the evaluator(s) to run. Use `None` if you only want to benchmark,
125
+ but don't want to do any evaluation.
126
+
127
+ Returns:
128
+ The return value of `evaluator.evaluate()`
129
+ """
130
+ num_devices = get_world_size()
131
+ logger = logging.getLogger(__name__)
132
+ logger.info("Start inference on {} batches".format(len(data_loader)))
133
+
134
+ total = len(data_loader) # inference data loader must have a fixed length
135
+ if evaluator is None:
136
+ # create a no-op evaluator
137
+ evaluator = DatasetEvaluators([])
138
+ if isinstance(evaluator, abc.MutableSequence):
139
+ evaluator = DatasetEvaluators(evaluator)
140
+ evaluator.reset()
141
+
142
+ num_warmup = min(5, total - 1)
143
+ start_time = time.perf_counter()
144
+ total_data_time = 0
145
+ total_compute_time = 0
146
+ total_eval_time = 0
147
+ with ExitStack() as stack:
148
+ if isinstance(model, nn.Module):
149
+ stack.enter_context(inference_context(model))
150
+ stack.enter_context(torch.no_grad())
151
+
152
+ start_data_time = time.perf_counter()
153
+ for idx, inputs in enumerate(data_loader):
154
+ total_data_time += time.perf_counter() - start_data_time
155
+ if idx == num_warmup:
156
+ start_time = time.perf_counter()
157
+ total_data_time = 0
158
+ total_compute_time = 0
159
+ total_eval_time = 0
160
+
161
+ start_compute_time = time.perf_counter()
162
+ outputs = model(inputs)
163
+ if torch.cuda.is_available():
164
+ torch.cuda.synchronize()
165
+ total_compute_time += time.perf_counter() - start_compute_time
166
+
167
+ start_eval_time = time.perf_counter()
168
+ evaluator.process(inputs, outputs)
169
+ total_eval_time += time.perf_counter() - start_eval_time
170
+
171
+ iters_after_start = idx + 1 - num_warmup * int(idx >= num_warmup)
172
+ data_seconds_per_iter = total_data_time / iters_after_start
173
+ compute_seconds_per_iter = total_compute_time / iters_after_start
174
+ eval_seconds_per_iter = total_eval_time / iters_after_start
175
+ total_seconds_per_iter = (time.perf_counter() - start_time) / iters_after_start
176
+ if idx >= num_warmup * 2 or compute_seconds_per_iter > 5:
177
+ eta = datetime.timedelta(seconds=int(total_seconds_per_iter * (total - idx - 1)))
178
+ log_every_n_seconds(
179
+ logging.INFO,
180
+ (
181
+ f"Inference done {idx + 1}/{total}. "
182
+ f"Dataloading: {data_seconds_per_iter:.4f} s/iter. "
183
+ f"Inference: {compute_seconds_per_iter:.4f} s/iter. "
184
+ f"Eval: {eval_seconds_per_iter:.4f} s/iter. "
185
+ f"Total: {total_seconds_per_iter:.4f} s/iter. "
186
+ f"ETA={eta}"
187
+ ),
188
+ n=5,
189
+ )
190
+ start_data_time = time.perf_counter()
191
+
192
+ # Measure the time only for this worker (before the synchronization barrier)
193
+ total_time = time.perf_counter() - start_time
194
+ total_time_str = str(datetime.timedelta(seconds=total_time))
195
+ # NOTE this format is parsed by grep
196
+ logger.info(
197
+ "Total inference time: {} ({:.6f} s / iter per device, on {} devices)".format(
198
+ total_time_str, total_time / (total - num_warmup), num_devices
199
+ )
200
+ )
201
+ total_compute_time_str = str(datetime.timedelta(seconds=int(total_compute_time)))
202
+ logger.info(
203
+ "Total inference pure compute time: {} ({:.6f} s / iter per device, on {} devices)".format(
204
+ total_compute_time_str, total_compute_time / (total - num_warmup), num_devices
205
+ )
206
+ )
207
+
208
+ results = evaluator.evaluate()
209
+ # An evaluator may return None when not in main process.
210
+ # Replace it by an empty dict instead to make it easier for downstream code to handle
211
+ if results is None:
212
+ results = {}
213
+ return results
214
+
215
+
216
+ @contextmanager
217
+ def inference_context(model):
218
+ """
219
+ A context where the model is temporarily changed to eval mode,
220
+ and restored to previous mode afterwards.
221
+
222
+ Args:
223
+ model: a torch Module
224
+ """
225
+ training_mode = model.training
226
+ model.eval()
227
+ yield
228
+ model.train(training_mode)
oneformer/evaluation/instance_evaluation.py ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ------------------------------------------------------------------------------
2
+ # Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/evaluation/instance_evaluation.py
3
+ # ------------------------------------------------------------------------------
4
+
5
+ import contextlib
6
+ import copy
7
+ import io
8
+ import itertools
9
+ import json
10
+ import logging
11
+ import numpy as np
12
+ import os
13
+ import pickle
14
+ from collections import OrderedDict
15
+ import pycocotools.mask as mask_util
16
+ import torch
17
+ from pycocotools.coco import COCO
18
+ from pycocotools.cocoeval import COCOeval
19
+ from tabulate import tabulate
20
+
21
+ import detectron2.utils.comm as comm
22
+ from detectron2.config import CfgNode
23
+ from detectron2.data import MetadataCatalog
24
+ from detectron2.data.datasets.coco import convert_to_coco_json
25
+ from detectron2.evaluation.coco_evaluation import COCOEvaluator, _evaluate_predictions_on_coco
26
+ from detectron2.evaluation.fast_eval_api import COCOeval_opt
27
+ from detectron2.structures import Boxes, BoxMode, pairwise_iou
28
+ from detectron2.utils.file_io import PathManager
29
+ from detectron2.utils.logger import create_small_table
30
+
31
+
32
+ # modified from COCOEvaluator for instance segmetnat
33
+ class InstanceSegEvaluator(COCOEvaluator):
34
+ """
35
+ Evaluate AR for object proposals, AP for instance detection/segmentation, AP
36
+ for keypoint detection outputs using COCO's metrics.
37
+ See http://cocodataset.org/#detection-eval and
38
+ http://cocodataset.org/#keypoints-eval to understand its metrics.
39
+ The metrics range from 0 to 100 (instead of 0 to 1), where a -1 or NaN means
40
+ the metric cannot be computed (e.g. due to no predictions made).
41
+
42
+ In addition to COCO, this evaluator is able to support any bounding box detection,
43
+ instance segmentation, or keypoint detection dataset.
44
+ """
45
+
46
+ def _eval_predictions(self, predictions, img_ids=None):
47
+ """
48
+ Evaluate predictions. Fill self._results with the metrics of the tasks.
49
+ """
50
+ self._logger.info("Preparing results for COCO format ...")
51
+ coco_results = list(itertools.chain(*[x["instances"] for x in predictions]))
52
+ tasks = self._tasks or self._tasks_from_predictions(coco_results)
53
+
54
+ # unmap the category ids for COCO
55
+ if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
56
+ dataset_id_to_contiguous_id = self._metadata.thing_dataset_id_to_contiguous_id
57
+ # all_contiguous_ids = list(dataset_id_to_contiguous_id.values())
58
+ # num_classes = len(all_contiguous_ids)
59
+ # assert min(all_contiguous_ids) == 0 and max(all_contiguous_ids) == num_classes - 1
60
+
61
+ reverse_id_mapping = {v: k for k, v in dataset_id_to_contiguous_id.items()}
62
+ for result in coco_results:
63
+ category_id = result["category_id"]
64
+ # assert category_id < num_classes, (
65
+ # f"A prediction has class={category_id}, "
66
+ # f"but the dataset only has {num_classes} classes and "
67
+ # f"predicted class id should be in [0, {num_classes - 1}]."
68
+ # )
69
+ assert category_id in reverse_id_mapping, (
70
+ f"A prediction has class={category_id}, "
71
+ f"but the dataset only has class ids in {dataset_id_to_contiguous_id}."
72
+ )
73
+ result["category_id"] = reverse_id_mapping[category_id]
74
+
75
+ if self._output_dir:
76
+ file_path = os.path.join(self._output_dir, "coco_instances_results.json")
77
+ self._logger.info("Saving results to {}".format(file_path))
78
+ with PathManager.open(file_path, "w") as f:
79
+ f.write(json.dumps(coco_results))
80
+ f.flush()
81
+
82
+ if not self._do_evaluation:
83
+ self._logger.info("Annotations are not available for evaluation.")
84
+ return
85
+
86
+ self._logger.info(
87
+ "Evaluating predictions with {} COCO API...".format(
88
+ "unofficial" if self._use_fast_impl else "official"
89
+ )
90
+ )
91
+ for task in sorted(tasks):
92
+ assert task in {"bbox", "segm", "keypoints"}, f"Got unknown task: {task}!"
93
+ coco_eval = (
94
+ _evaluate_predictions_on_coco(
95
+ self._coco_api,
96
+ coco_results,
97
+ task,
98
+ kpt_oks_sigmas=self._kpt_oks_sigmas,
99
+ use_fast_impl=self._use_fast_impl,
100
+ img_ids=img_ids,
101
+ max_dets_per_image=self._max_dets_per_image,
102
+ )
103
+ if len(coco_results) > 0
104
+ else None # cocoapi does not handle empty results very well
105
+ )
106
+
107
+ res = self._derive_coco_results(
108
+ coco_eval, task, class_names=self._metadata.get("thing_classes")
109
+ )
110
+ self._results[task] = res
oneformer/modeling/.DS_Store ADDED
Binary file (6.15 kB). View file