add: efficientloftr

hloc/match_dense.py

@@ -44,6 +44,25 @@ confs = {
         "max_error": 1,  # max error for assigned keypoints (in px)
         "cell_size": 1,  # size of quantization patch (max 1 kp/patch)
     },
+    "eloftr": {
+        "output": "matches-eloftr",
+        "model": {
+            "name": "eloftr",
+            "weights": "weights/eloftr_outdoor.ckpt",
+            "max_keypoints": 2000,
+            "match_threshold": 0.2,
+        },
+        "preprocessing": {
+            "grayscale": True,
+            "resize_max": 1024,
+            "dfactor": 32,
+            "width": 640,
+            "height": 480,
+            "force_resize": True,
+        },
+        "max_error": 1,  # max error for assigned keypoints (in px)
+        "cell_size": 1,  # size of quantization patch (max 1 kp/patch)
+    },
     # "loftr_quadtree": {
     #     "output": "matches-loftr-quadtree",
     #     "model": {

hloc/matchers/eloftr.py

@@ -0,0 +1,88 @@
+import sys
+import warnings
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+
+eloftr_path = Path(__file__).parent / "../../third_party/EfficientLoFTR"
+sys.path.append(str(eloftr_path))
+
+from src.loftr import LoFTR as ELoFTR_
+from src.loftr import full_default_cfg, opt_default_cfg, reparameter
+
+from hloc import logger
+
+from ..utils.base_model import BaseModel
+
+
+class LoFTR(BaseModel):
+    default_conf = {
+        "weights": "weights/eloftr_outdoor.ckpt",
+        "match_threshold": 0.2,
+        # "sinkhorn_iterations": 20,
+        "max_keypoints": -1,
+        # You can choose model type in ['full', 'opt']
+        "model_type": "full",  # 'full' for best quality, 'opt' for best efficiency
+        # You can choose numerical precision in ['fp32', 'mp', 'fp16']. 'fp16' for best efficiency
+        "precision": "fp32",
+    }
+    required_inputs = ["image0", "image1"]
+
+    def _init(self, conf):
+
+        if self.conf["model_type"] == "full":
+            _default_cfg = deepcopy(full_default_cfg)
+        elif self.conf["model_type"] == "opt":
+            _default_cfg = deepcopy(opt_default_cfg)
+
+        if self.conf["precision"] == "mp":
+            _default_cfg["mp"] = True
+        elif self.conf["precision"] == "fp16":
+            _default_cfg["half"] = True
+        model_path = eloftr_path / self.conf["weights"]
+        cfg = _default_cfg
+        cfg["match_coarse"]["thr"] = conf["match_threshold"]
+        # cfg["match_coarse"]["skh_iters"] = conf["sinkhorn_iterations"]
+        state_dict = torch.load(model_path, map_location="cpu")["state_dict"]
+        matcher = ELoFTR_(config=cfg)
+        matcher.load_state_dict(state_dict)
+        self.net = reparameter(matcher)
+
+        if self.conf["precision"] == "fp16":
+            self.net = self.net.half()
+        logger.info(f"Loaded Efficient LoFTR with weights {conf['weights']}")
+
+    def _forward(self, data):
+        # For consistency with hloc pairs, we refine kpts in image0!
+        rename = {
+            "keypoints0": "keypoints1",
+            "keypoints1": "keypoints0",
+            "image0": "image1",
+            "image1": "image0",
+            "mask0": "mask1",
+            "mask1": "mask0",
+        }
+        data_ = {rename[k]: v for k, v in data.items()}
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            pred = self.net(data_)
+        pred = {
+            "keypoints0": data_["mkpts0_f"],
+            "keypoints1": data_["mkpts1_f"],
+        }
+        scores = data_["mconf"]
+
+        top_k = self.conf["max_keypoints"]
+        if top_k is not None and len(scores) > top_k:
+            keep = torch.argsort(scores, descending=True)[:top_k]
+            pred["keypoints0"], pred["keypoints1"] = (
+                pred["keypoints0"][keep],
+                pred["keypoints1"][keep],
+            )
+            scores = scores[keep]
+
+        # Switch back indices
+        pred = {(rename[k] if k in rename else k): v for k, v in pred.items()}
+        pred["scores"] = scores
+        return pred

third_party/EfficientLoFTR/.gitignore

@@ -0,0 +1,12 @@
+.vscode/
+__pycache__/
+*.pyc
+*.DS_Store
+*.swp
+*.pth
+tmp.*
+*/.ipynb_checkpoints/*
+
+logs/
+weights/
+dump/

third_party/EfficientLoFTR/README.md

@@ -0,0 +1,69 @@
+# Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed
+
+### [Project Page](https://zju3dv.github.io/efficientloftr) | [Paper](https://zju3dv.github.io/efficientloftr/files/EfficientLoFTR.pdf) 
+<br/>
+
+> Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed  
+> [Yifan Wang](https://github.com/wyf2020)<sup>\*</sup>, [Xingyi He](https://github.com/hxy-123)<sup>\*</sup>, [Sida Peng](https://pengsida.net), [Dongli Tan](https://github.com/Cuistiano), [Xiaowei Zhou](http://xzhou.me)  
+> CVPR 2024
+
+https://github.com/zju3dv/EfficientLoFTR/assets/69951260/40890d21-180e-4e70-aeba-219178b0d824
+
+## TODO List
+- [x] Inference code and pretrained models
+- [x] Code for reproducing the test-set results
+- [ ] Add options of flash-attention and torch.compiler for better performance
+- [x] jupyter notebook demo for matching a pair of images
+- [ ] Training code
+
+## Installation
+```shell
+conda env create -f environment.yaml
+conda activate eloftr
+pip install torch==2.0.0+cu118 --index-url https://download.pytorch.org/whl/cu118
+pip install -r requirements.txt 
+```
+The test and training can be downloaded by [download link](https://drive.google.com/drive/folders/1DOcOPZb3-5cWxLqn256AhwUVjBPifhuf?usp=sharing) provided by LoFTR
+
+We provide the our pretrained model in [download link](https://drive.google.com/drive/folders/1GOw6iVqsB-f1vmG6rNmdCcgwfB4VZ7_Q?usp=sharing)
+
+
+## Reproduce the testing results with pytorch-lightning
+You need to setup the testing subsets of ScanNet and MegaDepth first. We create symlinks from the previously downloaded datasets to `data/{{dataset}}/test`.
+
+```shell
+# set up symlinks
+ln -s /path/to/scannet-1500-testset/* /path/to/EfficientLoFTR/data/scannet/test
+ln -s /path/to/megadepth-1500-testset/* /path/to/EfficientLoFTR/data/megadepth/test
+```
+### Inference time
+```shell
+conda activate eloftr
+bash scripts/reproduce_test/indoor_full_time.sh
+bash scripts/reproduce_test/indoor_opt_time.sh
+```
+
+### Accuracy
+```shell
+conda activate eloftr
+bash scripts/reproduce_test/outdoor_full_auc.sh
+bash scripts/reproduce_test/outdoor_opt_auc.sh
+bash scripts/reproduce_test/indoor_full_auc.sh
+bash scripts/reproduce_test/indoor_opt_auc.sh
+```
+
+## Training
+The Training code is coming soon, please stay tuned!
+
+## Citation
+
+If you find this code useful for your research, please use the following BibTeX entry.
+
+```bibtex
+@inproceedings{wang2024eloftr,
+  title={{Efficient LoFTR}: Semi-Dense Local Feature Matching with Sparse-Like Speed},
+  author={Wang, Yifan and He, Xingyi and Peng, Sida and Tan, Dongli and Zhou, Xiaowei},
+  booktitle={CVPR},
+  year={2024}
+}
+```

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/0015_0.1_0.3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d441df1d380b2ed34449b944d9f13127e695542fa275098d38a6298835672f22
+size 231253

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/0015_0.3_0.5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f34b5231d04a84d84378c671dd26854869663b5eafeae2ebaf624a279325139
+size 231253

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/0022_0.1_0.3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba46e6b9ec291fc7271eb9741d5c75ca04b83d3d7281e049815de9cb9024f4d9
+size 272610

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/0022_0.3_0.5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f4465da174b96deba61e5328886e4f2e687d34b890efca69e0c838736f8ae12
+size 272610

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/0022_0.5_0.7.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:684ae10f03001917c3ca0d12d441f372ce3c7e6637bd1277a3cda60df4207fe9
+size 272610

third_party/EfficientLoFTR/assets/megadepth_test_1500_scene_info/megadepth_test_1500.txt

@@ -0,0 +1,5 @@
+0022_0.1_0.3
+0015_0.1_0.3
+0015_0.3_0.5
+0022_0.3_0.5
+0022_0.5_0.7

third_party/EfficientLoFTR/assets/scannet_test_1500/intrinsics.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25ac102c69e2e4e2f0ab9c0d64f4da2b815e0901630768bdfde30080ced3605c
+size 23922

third_party/EfficientLoFTR/assets/scannet_test_1500/scannet_test.txt

@@ -0,0 +1 @@
+test.npz

third_party/EfficientLoFTR/assets/scannet_test_1500/statistics.json

@@ -0,0 +1,102 @@
+{
+    "scene0707_00": 15,
+    "scene0708_00": 15,
+    "scene0709_00": 15,
+    "scene0710_00": 15,
+    "scene0711_00": 15,
+    "scene0712_00": 15,
+    "scene0713_00": 15,
+    "scene0714_00": 15,
+    "scene0715_00": 15,
+    "scene0716_00": 15,
+    "scene0717_00": 15,
+    "scene0718_00": 15,
+    "scene0719_00": 15,
+    "scene0720_00": 15,
+    "scene0721_00": 15,
+    "scene0722_00": 15,
+    "scene0723_00": 15,
+    "scene0724_00": 15,
+    "scene0725_00": 15,
+    "scene0726_00": 15,
+    "scene0727_00": 15,
+    "scene0728_00": 15,
+    "scene0729_00": 15,
+    "scene0730_00": 15,
+    "scene0731_00": 15,
+    "scene0732_00": 15,
+    "scene0733_00": 15,
+    "scene0734_00": 15,
+    "scene0735_00": 15,
+    "scene0736_00": 15,
+    "scene0737_00": 15,
+    "scene0738_00": 15,
+    "scene0739_00": 15,
+    "scene0740_00": 15,
+    "scene0741_00": 15,
+    "scene0742_00": 15,
+    "scene0743_00": 15,
+    "scene0744_00": 15,
+    "scene0745_00": 15,
+    "scene0746_00": 15,
+    "scene0747_00": 15,
+    "scene0748_00": 15,
+    "scene0749_00": 15,
+    "scene0750_00": 15,
+    "scene0751_00": 15,
+    "scene0752_00": 15,
+    "scene0753_00": 15,
+    "scene0754_00": 15,
+    "scene0755_00": 15,
+    "scene0756_00": 15,
+    "scene0757_00": 15,
+    "scene0758_00": 15,
+    "scene0759_00": 15,
+    "scene0760_00": 15,
+    "scene0761_00": 15,
+    "scene0762_00": 15,
+    "scene0763_00": 15,
+    "scene0764_00": 15,
+    "scene0765_00": 15,
+    "scene0766_00": 15,
+    "scene0767_00": 15,
+    "scene0768_00": 15,
+    "scene0769_00": 15,
+    "scene0770_00": 15,
+    "scene0771_00": 15,
+    "scene0772_00": 15,
+    "scene0773_00": 15,
+    "scene0774_00": 15,
+    "scene0775_00": 15,
+    "scene0776_00": 15,
+    "scene0777_00": 15,
+    "scene0778_00": 15,
+    "scene0779_00": 15,
+    "scene0780_00": 15,
+    "scene0781_00": 15,
+    "scene0782_00": 15,
+    "scene0783_00": 15,
+    "scene0784_00": 15,
+    "scene0785_00": 15,
+    "scene0786_00": 15,
+    "scene0787_00": 15,
+    "scene0788_00": 15,
+    "scene0789_00": 15,
+    "scene0790_00": 15,
+    "scene0791_00": 15,
+    "scene0792_00": 15,
+    "scene0793_00": 15,
+    "scene0794_00": 15,
+    "scene0795_00": 15,
+    "scene0796_00": 15,
+    "scene0797_00": 15,
+    "scene0798_00": 15,
+    "scene0799_00": 15,
+    "scene0800_00": 15,
+    "scene0801_00": 15,
+    "scene0802_00": 15,
+    "scene0803_00": 15,
+    "scene0804_00": 15,
+    "scene0805_00": 15,
+    "scene0806_00": 15
+}

third_party/EfficientLoFTR/assets/scannet_test_1500/test.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b982b9c1f762e7d31af552ecc1ccf1a6add013197f74ec69c84a6deaa6f580ad
+size 71687

third_party/EfficientLoFTR/configs/data/base.py

@@ -0,0 +1,35 @@
+"""
+The data config will be the last one merged into the main config.
+Setups in data configs will override all existed setups!
+"""
+
+from yacs.config import CfgNode as CN
+_CN = CN()
+_CN.DATASET = CN()
+_CN.TRAINER = CN()
+
+# training data config
+_CN.DATASET.TRAIN_DATA_ROOT = None
+_CN.DATASET.TRAIN_POSE_ROOT = None
+_CN.DATASET.TRAIN_NPZ_ROOT = None
+_CN.DATASET.TRAIN_LIST_PATH = None
+_CN.DATASET.TRAIN_INTRINSIC_PATH = None
+# validation set config
+_CN.DATASET.VAL_DATA_ROOT = None
+_CN.DATASET.VAL_POSE_ROOT = None
+_CN.DATASET.VAL_NPZ_ROOT = None
+_CN.DATASET.VAL_LIST_PATH = None
+_CN.DATASET.VAL_INTRINSIC_PATH = None
+
+# testing data config
+_CN.DATASET.TEST_DATA_ROOT = None
+_CN.DATASET.TEST_POSE_ROOT = None
+_CN.DATASET.TEST_NPZ_ROOT = None
+_CN.DATASET.TEST_LIST_PATH = None
+_CN.DATASET.TEST_INTRINSIC_PATH = None
+
+# dataset config
+_CN.DATASET.MIN_OVERLAP_SCORE_TRAIN = 0.4
+_CN.DATASET.MIN_OVERLAP_SCORE_TEST = 0.0  # for both test and val
+
+cfg = _CN

third_party/EfficientLoFTR/configs/data/debug/.gitignore

@@ -0,0 +1,3 @@
+*
+*/
+!.gitignore

third_party/EfficientLoFTR/configs/data/megadepth_test_1500.py

@@ -0,0 +1,13 @@
+from configs.data.base import cfg
+
+TEST_BASE_PATH = "assets/megadepth_test_1500_scene_info"
+
+cfg.DATASET.TEST_DATA_SOURCE = "MegaDepth"
+cfg.DATASET.TEST_DATA_ROOT = "data/megadepth/test"
+cfg.DATASET.TEST_NPZ_ROOT = f"{TEST_BASE_PATH}"
+cfg.DATASET.TEST_LIST_PATH = f"{TEST_BASE_PATH}/megadepth_test_1500.txt"
+
+cfg.DATASET.MGDPT_IMG_RESIZE = 832
+cfg.DATASET.MIN_OVERLAP_SCORE_TEST = 0.0
+
+cfg.DATASET.NPE_NAME = 'megadepth'

third_party/EfficientLoFTR/configs/data/megadepth_trainval_832.py

@@ -0,0 +1,24 @@
+from configs.data.base import cfg
+
+
+TRAIN_BASE_PATH = "data/megadepth/index"
+cfg.DATASET.TRAINVAL_DATA_SOURCE = "MegaDepth"
+cfg.DATASET.TRAIN_DATA_ROOT = "data/megadepth/train"
+cfg.DATASET.TRAIN_NPZ_ROOT = f"{TRAIN_BASE_PATH}/scene_info_0.1_0.7"
+cfg.DATASET.TRAIN_LIST_PATH = f"{TRAIN_BASE_PATH}/trainvaltest_list/train_list.txt"
+cfg.DATASET.MIN_OVERLAP_SCORE_TRAIN = 0.0
+
+TEST_BASE_PATH = "data/megadepth/index"
+cfg.DATASET.TEST_DATA_SOURCE = "MegaDepth"
+cfg.DATASET.VAL_DATA_ROOT = cfg.DATASET.TEST_DATA_ROOT = "data/megadepth/test"
+cfg.DATASET.VAL_NPZ_ROOT = cfg.DATASET.TEST_NPZ_ROOT = f"{TEST_BASE_PATH}/scene_info_val_1500"
+cfg.DATASET.VAL_LIST_PATH = cfg.DATASET.TEST_LIST_PATH = f"{TEST_BASE_PATH}/trainvaltest_list/val_list.txt"
+cfg.DATASET.MIN_OVERLAP_SCORE_TEST = 0.0   # for both test and val
+
+# 368 scenes in total for MegaDepth
+# (with difficulty balanced (further split each scene to 3 sub-scenes))
+cfg.TRAINER.N_SAMPLES_PER_SUBSET = 100
+
+cfg.DATASET.MGDPT_IMG_RESIZE = 832  # for training on 32GB meme GPUs
+
+cfg.DATASET.NPE_NAME = 'megadepth'

third_party/EfficientLoFTR/configs/data/scannet_test_1500.py

@@ -0,0 +1,16 @@
+from configs.data.base import cfg
+
+TEST_BASE_PATH = "assets/scannet_test_1500"
+
+cfg.DATASET.TEST_DATA_SOURCE = "ScanNet"
+cfg.DATASET.TEST_DATA_ROOT = "data/scannet/test"
+cfg.DATASET.TEST_NPZ_ROOT = f"{TEST_BASE_PATH}"
+cfg.DATASET.TEST_LIST_PATH = f"{TEST_BASE_PATH}/scannet_test.txt"
+cfg.DATASET.TEST_INTRINSIC_PATH = f"{TEST_BASE_PATH}/intrinsics.npz"
+
+cfg.DATASET.MIN_OVERLAP_SCORE_TEST = 0.0
+
+cfg.DATASET.SCAN_IMG_RESIZEX = 640
+cfg.DATASET.SCAN_IMG_RESIZEY = 480
+
+cfg.DATASET.NPE_NAME = 'scannet'

third_party/EfficientLoFTR/configs/loftr/eloftr_full.py

@@ -0,0 +1,36 @@
+from src.config.default import _CN as cfg
+
+# training config
+cfg.TRAINER.CANONICAL_LR = 8e-3
+cfg.TRAINER.WARMUP_STEP = 1875  # 3 epochs
+cfg.TRAINER.WARMUP_RATIO = 0.1
+cfg.TRAINER.MSLR_MILESTONES = [8, 12, 16, 20, 24]
+cfg.TRAINER.RANSAC_PIXEL_THR = 0.5
+cfg.TRAINER.OPTIMIZER = "adamw"
+cfg.TRAINER.ADAMW_DECAY = 0.1
+cfg.TRAINER.EPI_ERR_THR = 5e-4 # recommendation: 5e-4 for ScanNet, 1e-4 for MegaDepth (from SuperGlue)
+cfg.TRAINER.GRADIENT_CLIPPING = 0.0
+cfg.LOFTR.LOSS.FINE_TYPE = 'l2'  # ['l2_with_std', 'l2']
+cfg.LOFTR.LOSS.COARSE_OVERLAP_WEIGHT = True
+cfg.LOFTR.LOSS.FINE_OVERLAP_WEIGHT = True
+cfg.LOFTR.LOSS.LOCAL_WEIGHT = 0.25
+cfg.LOFTR.MATCH_COARSE.TRAIN_COARSE_PERCENT = 0.3
+cfg.LOFTR.MATCH_COARSE.SPARSE_SPVS = True
+
+# model config
+cfg.LOFTR.RESOLUTION = (8, 1)
+cfg.LOFTR.FINE_WINDOW_SIZE = 8  # window_size in fine_level, must be even
+cfg.LOFTR.ALIGN_CORNER = False
+cfg.LOFTR.MP = True # just for reproducing paper, FP16 is much faster on modern GPUs
+cfg.LOFTR.REPLACE_NAN = True
+cfg.LOFTR.EVAL_TIMES = 5
+cfg.LOFTR.COARSE.NO_FLASH = True # Not use Flash-Attention just for reproducing paper timing
+cfg.LOFTR.MATCH_COARSE.THR = 0.2 # recommend 0.2 for full model and 25 for optimized model
+cfg.LOFTR.MATCH_FINE.LOCAL_REGRESS_TEMPERATURE = 10.0
+cfg.LOFTR.MATCH_FINE.LOCAL_REGRESS_SLICEDIM = 8
+
+# dataset config
+cfg.DATASET.FP16 = False
+
+# full model config
+cfg.LOFTR.MATCH_COARSE.FP16MATMUL = False

third_party/EfficientLoFTR/configs/loftr/eloftr_optimized.py

@@ -0,0 +1,37 @@
+from src.config.default import _CN as cfg
+
+# training config
+cfg.TRAINER.CANONICAL_LR = 8e-3
+cfg.TRAINER.WARMUP_STEP = 1875  # 3 epochs
+cfg.TRAINER.WARMUP_RATIO = 0.1
+cfg.TRAINER.MSLR_MILESTONES = [8, 12, 16, 20, 24]
+cfg.TRAINER.RANSAC_PIXEL_THR = 0.5
+cfg.TRAINER.OPTIMIZER = "adamw"
+cfg.TRAINER.ADAMW_DECAY = 0.1
+cfg.TRAINER.EPI_ERR_THR = 5e-4 # recommendation: 5e-4 for ScanNet, 1e-4 for MegaDepth (from SuperGlue)
+cfg.TRAINER.GRADIENT_CLIPPING = 0.0
+cfg.LOFTR.LOSS.FINE_TYPE = 'l2'  # ['l2_with_std', 'l2']
+cfg.LOFTR.LOSS.COARSE_OVERLAP_WEIGHT = True
+cfg.LOFTR.LOSS.FINE_OVERLAP_WEIGHT = True
+cfg.LOFTR.LOSS.LOCAL_WEIGHT = 0.25
+cfg.LOFTR.MATCH_COARSE.TRAIN_COARSE_PERCENT = 0.3
+cfg.LOFTR.MATCH_COARSE.SPARSE_SPVS = True
+
+# model config
+cfg.LOFTR.RESOLUTION = (8, 1)
+cfg.LOFTR.FINE_WINDOW_SIZE = 8  # window_size in fine_level, must be even
+cfg.LOFTR.ALIGN_CORNER = False
+cfg.LOFTR.MP = True # just for reproducing paper, FP16 is much faster on modern GPUs
+cfg.LOFTR.REPLACE_NAN = True
+cfg.LOFTR.EVAL_TIMES = 5
+cfg.LOFTR.COARSE.NO_FLASH = True # Not use Flash-Attention just for reproducing paper timing
+cfg.LOFTR.MATCH_FINE.LOCAL_REGRESS_TEMPERATURE = 10.0
+cfg.LOFTR.MATCH_FINE.LOCAL_REGRESS_SLICEDIM = 8
+
+# dataset config
+cfg.DATASET.FP16 = False
+
+# optimized model config
+cfg.LOFTR.MATCH_COARSE.FP16MATMUL = True
+cfg.LOFTR.MATCH_COARSE.SKIP_SOFTMAX = True
+cfg.LOFTR.MATCH_COARSE.THR = 25.0 # recommend 0.2 for full model and 25 for optimized model

third_party/EfficientLoFTR/data/megadepth/index/.gitignore

@@ -0,0 +1,4 @@
+# Ignore everything in this directory
+*
+# Except this file
+!.gitignore

third_party/EfficientLoFTR/data/megadepth/test/.gitignore

@@ -0,0 +1,4 @@
+# Ignore everything in this directory
+*
+# Except this file
+!.gitignore

third_party/EfficientLoFTR/data/megadepth/train/.gitignore

@@ -0,0 +1,4 @@
+# Ignore everything in this directory
+*
+# Except this file
+!.gitignore

third_party/EfficientLoFTR/data/scannet/index/.gitignore

@@ -0,0 +1,3 @@
+*
+*/
+!.gitignore

third_party/EfficientLoFTR/data/scannet/test/.gitignore

@@ -0,0 +1,3 @@
+*
+*/
+!.gitignore

third_party/EfficientLoFTR/environment.yaml

@@ -0,0 +1,7 @@
+name: eloftr
+channels:
+  - pytorch
+  - nvidia
+dependencies:
+  - python=3.8
+  - pip

third_party/EfficientLoFTR/requirements.txt

@@ -0,0 +1,17 @@
+opencv_python==4.4.0.46
+albumentations==0.5.1 --no-binary=imgaug,albumentations
+ray>=1.0.1
+einops==0.3.0
+kornia==0.4.1
+loguru==0.5.3
+yacs>=0.1.8
+tqdm
+autopep8
+pylint
+ipython
+jupyterlab
+matplotlib
+h5py==3.1.0
+pytorch-lightning==1.3.5
+torchmetrics==0.6.0  # version problem: https://github.com/NVIDIA/DeepLearningExamples/issues/1113#issuecomment-1102969461
+joblib>=1.0.1

third_party/EfficientLoFTR/scripts/reproduce_test/indoor_full_auc.sh

@@ -0,0 +1,35 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_full.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_scannet"
+data_cfg_path="configs/data/scannet_test_1500.py"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --scannetX '640' \
+    --scannetY '480' \
+    --rmbd 0 \
+    --thr 0.1 \
+    --deter \
+    --ransac_times 5
+# Following the RoMa protocol, we repeat RANSAC 5 times to enhance robustness; however, this increases script runtime.

third_party/EfficientLoFTR/scripts/reproduce_test/indoor_full_time.sh

@@ -0,0 +1,33 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_full.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_scannet"
+data_cfg_path="configs/data/scannet_test_1500.py"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --scannetX '640' \
+    --scannetY '480' \
+    --rmbd 0 \
+    --thr 0.1 \
+    --ransac_times 1

third_party/EfficientLoFTR/scripts/reproduce_test/indoor_opt_auc.sh

@@ -0,0 +1,35 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_optimized.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_scannet"
+data_cfg_path="configs/data/scannet_test_1500.py"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --scannetX '640' \
+    --scannetY '480' \
+    --rmbd 1 \
+    --thr 20 \
+    --deter \
+    --ransac_times 5
+# Following the RoMa protocol, we repeat RANSAC 5 times to enhance robustness; however, this increases script runtime.

third_party/EfficientLoFTR/scripts/reproduce_test/indoor_opt_time.sh

@@ -0,0 +1,33 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_optimized.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_scannet"
+data_cfg_path="configs/data/scannet_test_1500.py"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --scannetX '640' \
+    --scannetY '480' \
+    --rmbd 1 \
+    --thr 20 \
+    --ransac_times 1

third_party/EfficientLoFTR/scripts/reproduce_test/outdoor_full_auc.sh

@@ -0,0 +1,35 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_full.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_megadepth"
+data_cfg_path="configs/data/megadepth_test_1500.py"
+size="1152"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --megasize $size \
+    --npe \
+    --thr 0.1 \
+    --deter \
+    --ransac_times 5
+# Following the RoMa protocol, we repeat RANSAC 5 times to enhance robustness; however, this increases script runtime.

third_party/EfficientLoFTR/scripts/reproduce_test/outdoor_opt_auc.sh

@@ -0,0 +1,35 @@
+#!/bin/bash -l
+SCRIPTPATH=$(dirname $(readlink -f "$0"))
+PROJECT_DIR="${SCRIPTPATH}/../../"
+
+export PYTHONPATH=$PROJECT_DIR:$PYTHONPATH
+cd $PROJECT_DIR
+
+main_cfg_path="configs/loftr/eloftr_optimized.py"
+
+profiler_name="inference"
+n_nodes=1  # mannually keep this the same with --nodes
+n_gpus_per_node=-1
+torch_num_workers=4
+batch_size=1  # per gpu
+
+ckpt_path="weights/eloftr_outdoor.ckpt"
+
+dump_dir="dump/eloftr_full_megadepth"
+data_cfg_path="configs/data/megadepth_test_1500.py"
+size="1152"
+python ./test.py \
+    ${data_cfg_path} \
+    ${main_cfg_path} \
+    --ckpt_path=${ckpt_path} \
+    --dump_dir=${dump_dir} \
+    --gpus=${n_gpus_per_node} --num_nodes=${n_nodes} --accelerator="ddp" \
+    --batch_size=${batch_size} --num_workers=${torch_num_workers}\
+    --profiler_name=${profiler_name} \
+    --benchmark \
+    --megasize $size \
+    --npe \
+    --thr 20 \
+    --deter \
+    --ransac_times 5
+# Following the RoMa protocol, we repeat RANSAC 5 times to enhance robustness; however, this increases script runtime.