Datasets:

ArXiv:
Tags:
art
License:
schirrmacher commited on
Commit
9c206a9
1 Parent(s): d30ccb8

Upload folder using huggingface_hub

Browse files
.gitignore ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ gt/*
2
+ im/*
3
+ out/*
4
+ training/*
5
+ validation/*
6
+ dataset/*
README.md CHANGED
@@ -1,33 +1,41 @@
1
  ---
2
  license: apache-2.0
3
  tags:
4
- - art
5
  pretty_name: Human Segmentation Dataset
6
  ---
 
7
  # Human Segmentation Dataset
8
 
9
  This dataset was created **for developing the best fully open-source background remover** of images with humans.
10
  The dataset was crafted with [LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse), a Stable Diffusion extension for generating transparent images.
11
 
 
 
 
12
  The resulting model will be similar to [RMBG-1.4](https://huggingface.co/briaai/RMBG-1.4), but with open training data/process and commercially free to use.
13
 
14
- It covers a diverse set of humans: various skin tones, clothes, hair styles etc.
15
- Since Stable Diffusion is not perfect, the dataset contains images with flaws. Still the dataset is good enough for training background remover models.
 
16
 
17
- The dataset contains transparent images of humans (`/humans`) which were randomly combined with Stable Diffusion created backgrounds.
18
- Then the groundtruth (`/gt`) for segmentation was computed based on the transparent images. The results are written to a training and validation dataset.
19
 
20
- I created 4.558 images with people which I have augmented to a set of 9.116 images for training and 2.549 for validation with diverse backgrounds. The backgrounds are optimized for art exhibitions because this is where I want to apply the background remover.
 
 
 
 
 
 
 
 
21
 
22
  # Support
23
 
24
  If you identify weaknesses in the data, please contact me.
25
- I had some trouble with this huge file upload on huggingface. If files are missing use: [GDrvie Download: 61.1 GB](https://drive.google.com/drive/folders/1K1lK6nSoaQ7PLta-bcfol3XSGZA1b9nt?usp=drive_link)
26
-
27
- # Examples
28
 
29
- ![](training/im/aiznxclmqmkvi_tmpzjukj8v6.png)
30
- ![](training/gt/aiznxclmqmkvi_tmpzjukj8v6.png)
31
 
32
- ![](training/im/adcoyjxgttjef_tmp3tb7atsz.png)
33
- ![](training/gt/adcoyjxgttjef_tmp3tb7atsz.png)
 
1
  ---
2
  license: apache-2.0
3
  tags:
4
+ - art
5
  pretty_name: Human Segmentation Dataset
6
  ---
7
+
8
  # Human Segmentation Dataset
9
 
10
  This dataset was created **for developing the best fully open-source background remover** of images with humans.
11
  The dataset was crafted with [LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse), a Stable Diffusion extension for generating transparent images.
12
 
13
+ The dataset covers a diverse set of humans: various skin tones, clothes, hair styles etc.
14
+ Since Stable Diffusion is not perfect, the dataset contains images with flaws. Still the dataset is good enough for training background remover models.
15
+
16
  The resulting model will be similar to [RMBG-1.4](https://huggingface.co/briaai/RMBG-1.4), but with open training data/process and commercially free to use.
17
 
18
+ I had some trouble with the Hugging Face file upload. You can find the data here: [GDrvie](https://drive.google.com/drive/folders/1K1lK6nSoaQ7PLta-bcfol3XSGZA1b9nt?usp=drive_link)
19
+
20
+ The dataset contains transparent images of humans (`/humans`) which are randomly combined with backgrounds (`/backgrounds`). Then the ground truth (`/gt`) for segmentation was computed based on the transparent images. The results are written to a training and validation dataset.
21
 
22
+ I created more than 5.000 images with people and more than 5.000 diverse backgrounds.
 
23
 
24
+ # Create Training Dataset
25
+
26
+ The following scripts created training and validation data. Adding to this data is augmented.
27
+
28
+ Notice: download the dataset from [GDrvie](https://drive.google.com/drive/folders/1K1lK6nSoaQ7PLta-bcfol3XSGZA1b9nt?usp=drive_link).
29
+
30
+ ```
31
+ ./create_dataset.sh
32
+ ```
33
 
34
  # Support
35
 
36
  If you identify weaknesses in the data, please contact me.
 
 
 
37
 
38
+ # Changelog
 
39
 
40
+ - Added more diverse backgrounds (natural landscapes, streets, houses)
41
+ - Added more close-up images
backgrounds/background01.png ADDED

Git LFS Details

  • SHA256: fa2ffb53cd35dd4db66942f6d4372ee910f27076fc758a23d9a86420322ae350
  • Pointer size: 132 Bytes
  • Size of remote file: 1.47 MB
backgrounds/background02.png ADDED

Git LFS Details

  • SHA256: 38989e670e3350c31f5390c4f43f1c4e36895d53537491f0e1c788e5c1332650
  • Pointer size: 132 Bytes
  • Size of remote file: 1.15 MB
backgrounds/background03.png ADDED

Git LFS Details

  • SHA256: 6f1f92577aff294fa0e7b41c76408f44b8cae8d8b65cdf6b2d3ad16b70c725ca
  • Pointer size: 132 Bytes
  • Size of remote file: 1.82 MB
create_dataset.sh ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ random_merge() {
4
+ local backgrounds_dir="backgrounds"
5
+ local overlays_dir="humans"
6
+
7
+ local image_path="$1"
8
+ local groundtruth_path="$2"
9
+
10
+ background=$(find "$backgrounds_dir" -type f | shuf -n 1)
11
+ overlay=$(find "$overlays_dir" -type f | shuf -n 1)
12
+ echo "Processing iteration $i: $overlay + $background"
13
+ python3 "util/merge_images.py" \
14
+ -b "$background" -o "$overlay" \
15
+ -gt "$groundtruth_path" -im "$image_path"
16
+ }
17
+
18
+ main() {
19
+ local max_iterations=2000
20
+ for ((i = 0 ; i <= $max_iterations ; i++)); do
21
+ # For quicker creation some parallelization
22
+ # Notice: last iteration if for validation set
23
+ random_merge dataset/training/im dataset/training/gt &
24
+ random_merge dataset/training/im dataset/training/gt &
25
+ random_merge dataset/training/im dataset/training/gt &
26
+ random_merge dataset/training/im dataset/training/gt &
27
+ random_merge dataset/training/im dataset/training/gt &
28
+ random_merge dataset/training/im dataset/training/gt &
29
+ random_merge dataset/training/im dataset/training/gt &
30
+ random_merge dataset/training/im dataset/training/gt &
31
+ random_merge dataset/training/im dataset/training/gt &
32
+ random_merge dataset/validation/im dataset/validation/g
33
+ done
34
+ }
35
+
36
+ main
humans/example01.png ADDED

Git LFS Details

  • SHA256: a6a37d9eb2d51f35215b59d5f12d686682dbe927633d2d3f0721b0806f442acd
  • Pointer size: 132 Bytes
  • Size of remote file: 1.32 MB
humans/example02.png ADDED

Git LFS Details

  • SHA256: 462f3e7b90006df791ee17879655c3bf08ecedc69a3c5ce618150140e962502d
  • Pointer size: 131 Bytes
  • Size of remote file: 983 kB
humans/example03.png ADDED

Git LFS Details

  • SHA256: 0552d71d10a12cb6c4abf4d802e96859073ff30b02e7249860cc2936f177a683
  • Pointer size: 132 Bytes
  • Size of remote file: 1.41 MB
requirements.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ albumentations==1.4.6
2
+ opencv-python==4.9.0.80
util/merge_images.py ADDED
@@ -0,0 +1,265 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import cv2
3
+ import argparse
4
+ import random
5
+ import string
6
+ import albumentations as A
7
+
8
+
9
+ def apply_scale_and_move(image):
10
+ transform = A.Compose(
11
+ [
12
+ A.HorizontalFlip(p=0.5),
13
+ A.ShiftScaleRotate(
14
+ shift_limit_x=(-0.3, 0.3),
15
+ shift_limit_y=(0.0, 0.2),
16
+ scale_limit=(1.0, 1.5),
17
+ border_mode=cv2.BORDER_CONSTANT,
18
+ rotate_limit=(-3, 3),
19
+ p=0.7,
20
+ ),
21
+ ]
22
+ )
23
+ return transform(image=image)["image"]
24
+
25
+
26
+ def apply_transform(image):
27
+ has_alpha = image.shape[2] == 4
28
+ if has_alpha:
29
+ alpha_channel = image[:, :, 3]
30
+ color_channels = image[:, :, :3]
31
+ else:
32
+ color_channels = image
33
+
34
+ # Define the transformation
35
+ transform = A.Compose(
36
+ [
37
+ A.RandomBrightnessContrast(
38
+ brightness_limit=(-0.1, 0.1), contrast_limit=(-0.4, 0), p=0.8
39
+ )
40
+ ]
41
+ )
42
+
43
+ # Apply the transformation only to the color channels
44
+ transformed = transform(image=color_channels)
45
+ transformed_image = transformed["image"]
46
+
47
+ # Merge the alpha channel back if it was separated
48
+ if has_alpha:
49
+ final_image = cv2.merge(
50
+ (
51
+ transformed_image[:, :, 0],
52
+ transformed_image[:, :, 1],
53
+ transformed_image[:, :, 2],
54
+ alpha_channel,
55
+ )
56
+ )
57
+ else:
58
+ final_image = transformed_image
59
+ return final_image
60
+
61
+
62
+ def apply_noise(image):
63
+ transform = A.Compose(
64
+ [
65
+ A.MotionBlur(blur_limit=(5, 11), p=1.0),
66
+ A.GaussNoise(var_limit=(10, 150), p=1.0),
67
+ A.RandomBrightnessContrast(
68
+ brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5
69
+ ),
70
+ A.RandomFog(
71
+ fog_coef_lower=0.05,
72
+ fog_coef_upper=0.2,
73
+ alpha_coef=0.08,
74
+ always_apply=False,
75
+ p=0.5,
76
+ ),
77
+ A.RandomShadow(
78
+ shadow_roi=(0, 0.5, 1, 1),
79
+ num_shadows_limit=(1, 2),
80
+ num_shadows_lower=None,
81
+ num_shadows_upper=None,
82
+ shadow_dimension=5,
83
+ always_apply=False,
84
+ p=0.5,
85
+ ),
86
+ A.RandomToneCurve(scale=0.1, always_apply=False, p=0.5),
87
+ ]
88
+ )
89
+ return transform(image=image)["image"]
90
+
91
+
92
+ def remove_alpha(image, alpha_threshold=200):
93
+
94
+ mask = image[:, :, 3] < alpha_threshold
95
+ image[mask] = [0, 0, 0, 0]
96
+
97
+ return image
98
+
99
+
100
+ def merge_images(
101
+ background_path, overlay_path, output_path, groundtruth_path, width, height
102
+ ):
103
+ letters = string.ascii_lowercase
104
+ random_string = "".join(random.choice(letters) for i in range(13))
105
+ file_name = random_string + "_" + os.path.basename(overlay_path)
106
+
107
+ # Read the background image and resize it to the specified dimensions
108
+ background = cv2.imread(background_path, cv2.IMREAD_COLOR)
109
+
110
+ height, width = background.shape[:2]
111
+
112
+ height = int(1.5 * height)
113
+ width = int(1.5 * width)
114
+
115
+ resized_background = cv2.resize(
116
+ background, (width, height), interpolation=cv2.INTER_AREA
117
+ )
118
+
119
+ # Read the overlay image with alpha channel
120
+ overlay = cv2.imread(overlay_path, cv2.IMREAD_UNCHANGED)
121
+
122
+ # Ensure overlay has an alpha channel
123
+ if overlay.shape[2] < 4:
124
+ raise Exception("Overlay image does not have an alpha channel.")
125
+
126
+ # Apply transformations to the overlay
127
+ overlay = expand_image_borders_rgba(overlay, width, height)
128
+ overlay = apply_scale_and_move(overlay)
129
+
130
+ # store ground truth
131
+ extract_alpha_channel_as_bw(overlay, os.path.join(groundtruth_path, file_name))
132
+
133
+ overlay = apply_transform(overlay)
134
+
135
+ # Overlay placement on the resized background
136
+ x_offset = (width - overlay.shape[1]) // 2
137
+ y_offset = (height - overlay.shape[0]) // 2
138
+
139
+ # Preventing overlay from exceeding the background dimensions
140
+ x_offset = max(0, x_offset)
141
+ y_offset = max(0, y_offset)
142
+
143
+ # Calculate the normalized alpha mask
144
+ alpha_overlay = overlay[..., 3] / 255.0
145
+ region_of_interest = resized_background[
146
+ y_offset : y_offset + overlay.shape[0],
147
+ x_offset : x_offset + overlay.shape[1],
148
+ :,
149
+ ]
150
+
151
+ # Blend the images
152
+ for c in range(0, 3):
153
+ region_of_interest[..., c] = (
154
+ alpha_overlay * overlay[..., c]
155
+ + (1 - alpha_overlay) * region_of_interest[..., c]
156
+ )
157
+
158
+ resized_background[
159
+ y_offset : y_offset + overlay.shape[0], x_offset : x_offset + overlay.shape[1]
160
+ ] = region_of_interest
161
+
162
+ resized_background = apply_noise(resized_background)
163
+
164
+ cv2.imwrite(os.path.join(output_path, file_name), resized_background)
165
+
166
+
167
+ def expand_image_borders_rgba(
168
+ image, final_width, final_height, border_color=(0, 0, 0, 0)
169
+ ):
170
+ # Check if image has an alpha channel
171
+ if image.shape[2] < 4:
172
+ raise ValueError(
173
+ "Loaded image does not contain an alpha channel. Make sure the input image is RGBA."
174
+ )
175
+
176
+ # Current dimensions
177
+ height, width = image.shape[:2]
178
+
179
+ # Calculate padding needed
180
+ top = bottom = (final_height - height) // 2
181
+ left = right = (final_width - width) // 2
182
+
183
+ # To handle cases where the new dimensions are odd and original dimensions are even (or vice versa)
184
+ if (final_height - height) % 2 != 0:
185
+ bottom += 1
186
+ if (final_width - width) % 2 != 0:
187
+ right += 1
188
+
189
+ # Apply make border with an RGBA color
190
+ new_image = cv2.copyMakeBorder(
191
+ image, top, bottom, left, right, cv2.BORDER_CONSTANT, value=border_color
192
+ )
193
+
194
+ return new_image
195
+
196
+
197
+ def extract_alpha_channel_as_bw(image, output_path):
198
+ # Check if the image has an alpha channel
199
+ if image.shape[2] < 4:
200
+ raise ValueError(
201
+ "Loaded image does not contain an alpha channel. Make sure the input image is in PNG format with an alpha channel."
202
+ )
203
+
204
+ # Extract the alpha channel
205
+ image = remove_alpha(image.copy())
206
+ alpha_channel = image[:, :, 3]
207
+ # Save or display the alpha channel as a black and white image
208
+ cv2.imwrite(output_path, alpha_channel)
209
+
210
+
211
+ def main():
212
+ parser = argparse.ArgumentParser(
213
+ description="Merge two images with one image having transparency."
214
+ )
215
+ parser.add_argument(
216
+ "-b", "--background", required=True, help="Path to the background image"
217
+ )
218
+ parser.add_argument(
219
+ "-o", "--overlay", required=True, help="Path to the overlay image"
220
+ )
221
+ parser.add_argument(
222
+ "-im",
223
+ "--image-path",
224
+ type=str,
225
+ default="im",
226
+ help="Path where the merged image will be saved",
227
+ )
228
+ parser.add_argument(
229
+ "--width",
230
+ type=int,
231
+ default=1920,
232
+ help="Width to which the background image will be resized",
233
+ )
234
+ parser.add_argument(
235
+ "--height",
236
+ type=int,
237
+ default=1080,
238
+ help="Height to which the background image will be resized",
239
+ )
240
+ parser.add_argument(
241
+ "-gt",
242
+ "--groundtruth-path",
243
+ type=str,
244
+ default="gt",
245
+ help="Ground truth folder",
246
+ )
247
+ args = parser.parse_args()
248
+
249
+ if not os.path.exists(args.image_path):
250
+ os.makedirs(args.image_path)
251
+ if not os.path.exists(args.groundtruth_path):
252
+ os.makedirs(args.groundtruth_path)
253
+
254
+ merge_images(
255
+ args.background,
256
+ args.overlay,
257
+ args.image_path,
258
+ args.groundtruth_path,
259
+ args.width,
260
+ args.height,
261
+ )
262
+
263
+
264
+ if __name__ == "__main__":
265
+ main()