Upload 4 files
Browse files- generate_occluded_imagenet.py +207 -0
- generate_occluded_imagenet_faster.py +139 -0
- generate_occluded_imagenet_single_occluder.py +156 -0
- occlusion_info.csv +0 -0
generate_occluded_imagenet.py
ADDED
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
from PIL import Image
|
6 |
+
from torchvision import datasets, transforms, io
|
7 |
+
|
8 |
+
|
9 |
+
def get_random_occluder(dataset_occluder):
|
10 |
+
index = random.randint(0, len(dataset_occluder) - 1)
|
11 |
+
texture_path, texture_class_index = dataset_occluder.imgs[index]
|
12 |
+
texture_class = dataset_occluder.classes[texture_class_index]
|
13 |
+
|
14 |
+
# load with the alpha channel
|
15 |
+
texture = io.read_image(texture_path, mode=io.image.ImageReadMode.RGB_ALPHA)
|
16 |
+
|
17 |
+
return texture, texture_class
|
18 |
+
|
19 |
+
|
20 |
+
def resize_occluder(occluder_pil, target_area):
|
21 |
+
alpha = np.array(occluder_pil.getchannel('A'))
|
22 |
+
non_transparent_area = np.count_nonzero(alpha > 0)
|
23 |
+
|
24 |
+
area_scale_factor = target_area / non_transparent_area
|
25 |
+
|
26 |
+
width_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.width / occluder_pil.height))
|
27 |
+
height_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.height / occluder_pil.width))
|
28 |
+
|
29 |
+
new_width = occluder_pil.width * width_scale_factor
|
30 |
+
new_height = occluder_pil.height * height_scale_factor
|
31 |
+
|
32 |
+
resized_occluder = occluder_pil.resize((int(new_width), int(new_height)), Image.LANCZOS)
|
33 |
+
|
34 |
+
return resized_occluder
|
35 |
+
|
36 |
+
|
37 |
+
def randomly_rotate_occluder(occluder_pil):
|
38 |
+
angle = random.uniform(-180, 180)
|
39 |
+
return occluder_pil.rotate(angle, resample=Image.BICUBIC, expand=True)
|
40 |
+
|
41 |
+
|
42 |
+
def calculate_distance(point1, point2):
|
43 |
+
x1, y1 = point1
|
44 |
+
x2, y2 = point2
|
45 |
+
return ((x2 - x1) ** 2 + (y2 - y1) ** 2) ** 0.5
|
46 |
+
|
47 |
+
|
48 |
+
def try_rotations(occluder_pil, image_pil, target_area, pos1=None):
|
49 |
+
best_occluder = None
|
50 |
+
best_area = 0
|
51 |
+
best_pos = None
|
52 |
+
min_distance = 130 # min distance if two occluders
|
53 |
+
for _ in range(75): # increase number of attempts to find better position
|
54 |
+
rotated = randomly_rotate_occluder(occluder_pil)
|
55 |
+
resized = resize_occluder(rotated, target_area)
|
56 |
+
if resized.width > image_pil.width or resized.height > image_pil.height:
|
57 |
+
pos = (image_pil.width // 2 - resized.width // 2,
|
58 |
+
image_pil.height // 2 - resized.height // 2)
|
59 |
+
else:
|
60 |
+
max_x = max(0, image_pil.width - resized.width)
|
61 |
+
max_y = max(0, image_pil.height - resized.height)
|
62 |
+
pos = (random.randint(0, max_x), random.randint(0, max_y))
|
63 |
+
|
64 |
+
if pos1 is not None and calculate_distance(pos1, pos) < min_distance:
|
65 |
+
continue
|
66 |
+
|
67 |
+
mask = Image.new('1', image_pil.size)
|
68 |
+
mask.paste(resized.getchannel('A'), pos, resized.getchannel('A'))
|
69 |
+
area = np.count_nonzero(np.array(mask))
|
70 |
+
|
71 |
+
if area > best_area:
|
72 |
+
best_area = area
|
73 |
+
best_occluder = resized
|
74 |
+
best_pos = pos
|
75 |
+
|
76 |
+
return best_occluder, best_pos
|
77 |
+
|
78 |
+
|
79 |
+
def occlude_image(image, occluder_tensor, percentage_occlusion, occluded_dir, img_name):
|
80 |
+
occluder_pil = transforms.ToPILImage(mode='RGBA')(occluder_tensor)
|
81 |
+
image_pil = transforms.ToPILImage()(image)
|
82 |
+
|
83 |
+
binary_mask = Image.new('1', image_pil.size)
|
84 |
+
|
85 |
+
rn = random.random()
|
86 |
+
|
87 |
+
if rn < 0.5: # randomly use two occluders
|
88 |
+
occluder_resizing_factor = 0.5
|
89 |
+
else:
|
90 |
+
occluder_resizing_factor = 1.0
|
91 |
+
|
92 |
+
target_area = image_pil.width * image_pil.height * percentage_occlusion * occluder_resizing_factor
|
93 |
+
|
94 |
+
if rn < 0.5: # randomly use two occluders (can make this k occluders)
|
95 |
+
occluder_pil1, pos1 = try_rotations(occluder_pil, image_pil, target_area / 2)
|
96 |
+
image_pil.paste(occluder_pil1, pos1, occluder_pil1)
|
97 |
+
occluder_alpha1 = occluder_pil1.getchannel('A')
|
98 |
+
binary_mask.paste(occluder_alpha1, pos1, occluder_alpha1)
|
99 |
+
|
100 |
+
occluder_pil2, pos2 = try_rotations(occluder_pil, image_pil, target_area / 2, pos1)
|
101 |
+
if occluder_pil2 is not None and pos2 is not None:
|
102 |
+
image_pil.paste(occluder_pil2, pos2, occluder_pil2)
|
103 |
+
occluder_alpha2 = occluder_pil2.getchannel('A')
|
104 |
+
binary_mask.paste(occluder_alpha2, pos2, occluder_alpha2)
|
105 |
+
|
106 |
+
if pos2 is None:
|
107 |
+
pos = [pos1]
|
108 |
+
else:
|
109 |
+
pos = [pos1, pos2]
|
110 |
+
else:
|
111 |
+
occluder_pil, pos = try_rotations(occluder_pil, image_pil, target_area)
|
112 |
+
image_pil.paste(occluder_pil, pos, occluder_pil)
|
113 |
+
occluder_alpha = occluder_pil.getchannel('A')
|
114 |
+
binary_mask.paste(occluder_alpha, pos, occluder_alpha)
|
115 |
+
|
116 |
+
pos = [pos]
|
117 |
+
|
118 |
+
image_with_occluder_tensor = transforms.ToTensor()(image_pil)
|
119 |
+
|
120 |
+
mask_array = np.array(binary_mask)
|
121 |
+
mask_path = os.path.join(occluded_dir, f"{img_name}_mask.npy")
|
122 |
+
np.save(mask_path, mask_array)
|
123 |
+
|
124 |
+
return image_with_occluder_tensor, mask_path, pos
|
125 |
+
|
126 |
+
|
127 |
+
def rebuild_display_mask(image_path, mask_path):
|
128 |
+
image_pil = Image.open(image_path)
|
129 |
+
binary_mask = Image.new('1', image_pil.size)
|
130 |
+
|
131 |
+
mask_array = np.load(mask_path)
|
132 |
+
mask_indices = np.transpose(np.nonzero(mask_array))
|
133 |
+
|
134 |
+
for i, j in mask_indices:
|
135 |
+
binary_mask.putpixel((j, i), 1)
|
136 |
+
|
137 |
+
binary_mask.show()
|
138 |
+
|
139 |
+
|
140 |
+
def build_dataset(data_path, transform):
|
141 |
+
dataset = datasets.ImageFolder(data_path, transform=transform)
|
142 |
+
nb_classes = len(dataset.classes)
|
143 |
+
return dataset, nb_classes
|
144 |
+
|
145 |
+
|
146 |
+
def build_transform():
|
147 |
+
t = []
|
148 |
+
t.append(transforms.ToTensor())
|
149 |
+
return transforms.Compose(t)
|
150 |
+
|
151 |
+
|
152 |
+
def main():
|
153 |
+
data_dir = 'imagenet'
|
154 |
+
texture_dir = 'occluders_segmented'
|
155 |
+
occluded_data_dir = 'imagenet_occluded'
|
156 |
+
|
157 |
+
transform = build_transform()
|
158 |
+
dataset, nb_classes = build_dataset(data_dir, transform)
|
159 |
+
dataset_occluder, _ = build_dataset(texture_dir, transform)
|
160 |
+
percentage_occlusion = 0.3 # ~30%
|
161 |
+
|
162 |
+
occlusion_info = pd.DataFrame(columns=["image_name", "class_name", "occluder_class",
|
163 |
+
"percentage_occlusion", "mask", "pos"])
|
164 |
+
|
165 |
+
count = 0
|
166 |
+
|
167 |
+
for idx in range(len(dataset)):
|
168 |
+
image, label = dataset[idx]
|
169 |
+
category = dataset.classes[label]
|
170 |
+
|
171 |
+
in_dir = os.path.join(data_dir, category)
|
172 |
+
occluded_dir = os.path.join(occluded_data_dir, category)
|
173 |
+
|
174 |
+
os.makedirs(occluded_dir, exist_ok=True)
|
175 |
+
|
176 |
+
img_name = dataset.imgs[idx][0].split('/')[-1].split('.')[0]
|
177 |
+
|
178 |
+
occluder_tensor, occluder_class = get_random_occluder(dataset_occluder)
|
179 |
+
occluded_image, mask_path, pos = occlude_image(image, occluder_tensor,
|
180 |
+
percentage_occlusion,
|
181 |
+
occluded_dir,
|
182 |
+
img_name)
|
183 |
+
|
184 |
+
mask_array = np.load(mask_path)
|
185 |
+
actual_percentage_occlusion = np.count_nonzero(mask_array) / (image.shape[1] * image.shape[2])
|
186 |
+
|
187 |
+
occluded_image_path = os.path.join(occluded_dir, f"{img_name}_occluded.png")
|
188 |
+
transforms.ToPILImage()(occluded_image).save(occluded_image_path)
|
189 |
+
|
190 |
+
new_row = pd.DataFrame({
|
191 |
+
"image_name": [f"{img_name}_occluded.png"],
|
192 |
+
"class_name": [category],
|
193 |
+
"occluder_class": [occluder_class],
|
194 |
+
"percentage_occlusion": [actual_percentage_occlusion],
|
195 |
+
"mask": [mask_path],
|
196 |
+
"pos": [pos]
|
197 |
+
})
|
198 |
+
|
199 |
+
occlusion_info = pd.concat([occlusion_info, new_row], ignore_index=True)
|
200 |
+
if count % 50 == 0:
|
201 |
+
print("Folder: {}/1000 ({})".format(count / 50, category))
|
202 |
+
count+=1
|
203 |
+
|
204 |
+
occlusion_info.to_csv(os.path.join(occluded_data_dir, "occlusion_info.csv"), index=False)
|
205 |
+
|
206 |
+
if __name__ == "__main__":
|
207 |
+
main()
|
generate_occluded_imagenet_faster.py
ADDED
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
from PIL import Image
|
6 |
+
from torchvision import datasets, transforms, io
|
7 |
+
import torch
|
8 |
+
|
9 |
+
|
10 |
+
def get_random_texture(dataset_occluder):
|
11 |
+
index = random.randint(0, len(dataset_occluder) - 1)
|
12 |
+
texture_path, texture_class_index = dataset_occluder.imgs[index]
|
13 |
+
texture_class = dataset_occluder.classes[texture_class_index]
|
14 |
+
|
15 |
+
# Load the texture with the alpha channel
|
16 |
+
texture = io.read_image(texture_path, mode=io.image.ImageReadMode.RGB_ALPHA)
|
17 |
+
|
18 |
+
return texture, texture_class
|
19 |
+
|
20 |
+
def resize_occluder(occluder_pil, target_area, image_width, image_height):
|
21 |
+
alpha = np.array(occluder_pil.getchannel('A'))
|
22 |
+
non_transparent_area = np.count_nonzero(alpha > 0)
|
23 |
+
|
24 |
+
area_scale_factor = target_area / non_transparent_area
|
25 |
+
|
26 |
+
width_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.width / occluder_pil.height))
|
27 |
+
height_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.height / occluder_pil.width))
|
28 |
+
|
29 |
+
new_width = occluder_pil.width * width_scale_factor
|
30 |
+
new_height = occluder_pil.height * height_scale_factor
|
31 |
+
|
32 |
+
resized_occluder = occluder_pil.resize((int(new_width), int(new_height)), Image.LANCZOS)
|
33 |
+
|
34 |
+
return resized_occluder
|
35 |
+
|
36 |
+
def randomly_rotate_occluder(occluder_pil):
|
37 |
+
angle = random.uniform(-180, 180)
|
38 |
+
return occluder_pil.rotate(angle, resample=Image.BICUBIC, expand=True)
|
39 |
+
|
40 |
+
def occlude_image(image, occluder_tensor, percentage_occlusion, occluded_dir, img_name):
|
41 |
+
occluder_pil = transforms.ToPILImage(mode='RGBA')(occluder_tensor)
|
42 |
+
occluder_pil = randomly_rotate_occluder(occluder_pil)
|
43 |
+
image_pil = transforms.ToPILImage()(image)
|
44 |
+
|
45 |
+
target_area = image_pil.width * image_pil.height * percentage_occlusion
|
46 |
+
occluder_pil = resize_occluder(occluder_pil, target_area, image_pil.width, image_pil.height)
|
47 |
+
|
48 |
+
if occluder_pil.width > image_pil.width or occluder_pil.height > image_pil.height:
|
49 |
+
pos = (image_pil.width // 2 - occluder_pil.width // 2,
|
50 |
+
image_pil.height // 2 - occluder_pil.height // 2)
|
51 |
+
else:
|
52 |
+
max_x = max(0, image_pil.width - occluder_pil.width)
|
53 |
+
max_y = max(0, image_pil.height - occluder_pil.height)
|
54 |
+
pos = (random.randint(0, max_x), random.randint(0, max_y))
|
55 |
+
|
56 |
+
image_pil.paste(occluder_pil, pos, occluder_pil)
|
57 |
+
image_with_occluder_tensor = transforms.ToTensor()(image_pil)
|
58 |
+
|
59 |
+
occluder_alpha = occluder_pil.getchannel('A')
|
60 |
+
binary_mask = Image.new('1', image_pil.size)
|
61 |
+
binary_mask.paste(occluder_alpha, pos, occluder_alpha)
|
62 |
+
|
63 |
+
mask_array = np.array(binary_mask)
|
64 |
+
mask_path = os.path.join(occluded_dir, f"{img_name}_mask.npy")
|
65 |
+
np.save(mask_path, mask_array)
|
66 |
+
|
67 |
+
return image_with_occluder_tensor, mask_path, pos
|
68 |
+
|
69 |
+
|
70 |
+
def rebuild_display_mask(image_path, mask_path):
|
71 |
+
image_pil = Image.open(image_path)
|
72 |
+
binary_mask = Image.new('1', image_pil.size)
|
73 |
+
|
74 |
+
mask_array = np.load(mask_path)
|
75 |
+
mask_indices = np.transpose(np.nonzero(mask_array))
|
76 |
+
|
77 |
+
for i, j in mask_indices:
|
78 |
+
binary_mask.putpixel((j, i), 1)
|
79 |
+
|
80 |
+
binary_mask.show()
|
81 |
+
|
82 |
+
|
83 |
+
def build_dataset(data_path, transform):
|
84 |
+
dataset = datasets.ImageFolder(data_path, transform=transform)
|
85 |
+
nb_classes = len(dataset.classes)
|
86 |
+
return dataset, nb_classes
|
87 |
+
|
88 |
+
def build_transform():
|
89 |
+
t = []
|
90 |
+
t.append(transforms.ToTensor())
|
91 |
+
return transforms.Compose(t)
|
92 |
+
|
93 |
+
def main():
|
94 |
+
data_dir = 'imagenet1'
|
95 |
+
texture_dir = 'occluders_segmented'
|
96 |
+
occluded_data_dir = 'imagenet_occluded'
|
97 |
+
|
98 |
+
transform = build_transform()
|
99 |
+
dataset, nb_classes = build_dataset(data_dir, transform)
|
100 |
+
dataset_occluder, _ = build_dataset(texture_dir, transform)
|
101 |
+
|
102 |
+
occlusion_info = pd.DataFrame(columns=["image_name", "class_name", "occluder_class",
|
103 |
+
"percentage_occlusion", "mask", "pos"])
|
104 |
+
|
105 |
+
for idx in range(len(dataset)):
|
106 |
+
image, label = dataset[idx]
|
107 |
+
category = dataset.classes[label]
|
108 |
+
|
109 |
+
in_dir = os.path.join(data_dir, category)
|
110 |
+
occluded_dir = os.path.join(occluded_data_dir, category)
|
111 |
+
|
112 |
+
os.makedirs(occluded_dir, exist_ok=True)
|
113 |
+
|
114 |
+
img_name = dataset.imgs[idx][0].split('/')[-1].split('.')[0]
|
115 |
+
|
116 |
+
occluder_tensor, occluder_class = get_random_texture(dataset_occluder)
|
117 |
+
occluded_image, mask_path, pos = occlude_image(image, occluder_tensor, 0.5, occluded_dir, img_name)
|
118 |
+
|
119 |
+
mask_array = np.load(mask_path)
|
120 |
+
actual_percentage_occlusion = np.count_nonzero(mask_array) / (image.shape[1] * image.shape[2])
|
121 |
+
|
122 |
+
occluded_image_path = os.path.join(occluded_dir, f"{img_name}_occluded.png")
|
123 |
+
transforms.ToPILImage()(occluded_image).save(occluded_image_path)
|
124 |
+
|
125 |
+
new_row = pd.DataFrame({
|
126 |
+
"image_name": [f"{img_name}_occluded.png"],
|
127 |
+
"class_name": [category],
|
128 |
+
"occluder_class": [occluder_class],
|
129 |
+
"percentage_occlusion": [actual_percentage_occlusion],
|
130 |
+
"mask": [mask_path],
|
131 |
+
"pos": [pos]
|
132 |
+
})
|
133 |
+
|
134 |
+
occlusion_info = pd.concat([occlusion_info, new_row], ignore_index=True)
|
135 |
+
|
136 |
+
occlusion_info.to_csv(os.path.join(occluded_data_dir, "occlusion_info.csv"), index=False)
|
137 |
+
|
138 |
+
if __name__ == "__main__":
|
139 |
+
main()
|
generate_occluded_imagenet_single_occluder.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
from PIL import Image
|
6 |
+
from torchvision import datasets, transforms, io
|
7 |
+
import torch
|
8 |
+
|
9 |
+
|
10 |
+
def get_random_texture(dataset_occluder):
|
11 |
+
index = random.randint(0, len(dataset_occluder) - 1)
|
12 |
+
texture_path, texture_class_index = dataset_occluder.imgs[index]
|
13 |
+
texture_class = dataset_occluder.classes[texture_class_index]
|
14 |
+
|
15 |
+
# Load the texture with the alpha channel
|
16 |
+
texture = io.read_image(texture_path, mode=io.image.ImageReadMode.RGB_ALPHA)
|
17 |
+
|
18 |
+
return texture, texture_class
|
19 |
+
|
20 |
+
def resize_occluder(occluder_pil, target_area, image_width, image_height):
|
21 |
+
alpha = np.array(occluder_pil.getchannel('A'))
|
22 |
+
non_transparent_area = np.count_nonzero(alpha > 0)
|
23 |
+
|
24 |
+
area_scale_factor = target_area / non_transparent_area
|
25 |
+
|
26 |
+
width_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.width / occluder_pil.height))
|
27 |
+
height_scale_factor = np.sqrt(area_scale_factor * (occluder_pil.height / occluder_pil.width))
|
28 |
+
|
29 |
+
new_width = occluder_pil.width * width_scale_factor
|
30 |
+
new_height = occluder_pil.height * height_scale_factor
|
31 |
+
|
32 |
+
resized_occluder = occluder_pil.resize((int(new_width), int(new_height)), Image.LANCZOS)
|
33 |
+
|
34 |
+
return resized_occluder
|
35 |
+
|
36 |
+
def randomly_rotate_occluder(occluder_pil):
|
37 |
+
angle = random.uniform(-180, 180)
|
38 |
+
return occluder_pil.rotate(angle, resample=Image.BICUBIC, expand=True)
|
39 |
+
|
40 |
+
def try_rotations(occluder_pil, image_pil, target_area):
|
41 |
+
best_occluder = None
|
42 |
+
best_area = 0
|
43 |
+
best_pos = None
|
44 |
+
for _ in range(10):
|
45 |
+
rotated = randomly_rotate_occluder(occluder_pil)
|
46 |
+
resized = resize_occluder(rotated, target_area, image_pil.width, image_pil.height)
|
47 |
+
if resized.width > image_pil.width or resized.height > image_pil.height:
|
48 |
+
pos = (image_pil.width // 2 - resized.width // 2,
|
49 |
+
image_pil.height // 2 - resized.height // 2)
|
50 |
+
else:
|
51 |
+
max_x = max(0, image_pil.width - resized.width)
|
52 |
+
max_y = max(0, image_pil.height - resized.height)
|
53 |
+
pos = (random.randint(0, max_x), random.randint(0, max_y))
|
54 |
+
|
55 |
+
mask = Image.new('1', image_pil.size)
|
56 |
+
mask.paste(resized.getchannel('A'), pos, resized.getchannel('A'))
|
57 |
+
area = np.count_nonzero(np.array(mask))
|
58 |
+
|
59 |
+
if area > best_area:
|
60 |
+
best_area = area
|
61 |
+
best_occluder = resized
|
62 |
+
best_pos = pos
|
63 |
+
return best_occluder, best_pos
|
64 |
+
|
65 |
+
|
66 |
+
def occlude_image(image, occluder_tensor, percentage_occlusion, occluded_dir, img_name):
|
67 |
+
occluder_pil = transforms.ToPILImage(mode='RGBA')(occluder_tensor)
|
68 |
+
image_pil = transforms.ToPILImage()(image)
|
69 |
+
|
70 |
+
target_area = image_pil.width * image_pil.height * percentage_occlusion
|
71 |
+
occluder_pil, pos = try_rotations(occluder_pil, image_pil, target_area)
|
72 |
+
|
73 |
+
image_pil.paste(occluder_pil, pos, occluder_pil)
|
74 |
+
image_with_occluder_tensor = transforms.ToTensor()(image_pil)
|
75 |
+
|
76 |
+
occluder_alpha = occluder_pil.getchannel('A')
|
77 |
+
binary_mask = Image.new('1', image_pil.size)
|
78 |
+
binary_mask.paste(occluder_alpha, pos, occluder_alpha)
|
79 |
+
|
80 |
+
mask_array = np.array(binary_mask)
|
81 |
+
mask_path = os.path.join(occluded_dir, f"{img_name}_mask.npy")
|
82 |
+
np.save(mask_path, mask_array)
|
83 |
+
|
84 |
+
return image_with_occluder_tensor, mask_path, pos
|
85 |
+
|
86 |
+
|
87 |
+
def rebuild_display_mask(image_path, mask_path):
|
88 |
+
image_pil = Image.open(image_path)
|
89 |
+
binary_mask = Image.new('1', image_pil.size)
|
90 |
+
|
91 |
+
mask_array = np.load(mask_path)
|
92 |
+
mask_indices = np.transpose(np.nonzero(mask_array))
|
93 |
+
|
94 |
+
for i, j in mask_indices:
|
95 |
+
binary_mask.putpixel((j, i), 1)
|
96 |
+
|
97 |
+
binary_mask.show()
|
98 |
+
|
99 |
+
|
100 |
+
def build_dataset(data_path, transform):
|
101 |
+
dataset = datasets.ImageFolder(data_path, transform=transform)
|
102 |
+
nb_classes = len(dataset.classes)
|
103 |
+
return dataset, nb_classes
|
104 |
+
|
105 |
+
def build_transform():
|
106 |
+
t = []
|
107 |
+
t.append(transforms.ToTensor())
|
108 |
+
return transforms.Compose(t)
|
109 |
+
|
110 |
+
def main():
|
111 |
+
data_dir = 'imagenet1'
|
112 |
+
texture_dir = 'occluders_segmented'
|
113 |
+
occluded_data_dir = 'imagenet_occluded'
|
114 |
+
|
115 |
+
transform = build_transform()
|
116 |
+
dataset, nb_classes = build_dataset(data_dir, transform)
|
117 |
+
dataset_occluder, _ = build_dataset(texture_dir, transform)
|
118 |
+
|
119 |
+
occlusion_info = pd.DataFrame(columns=["image_name", "class_name", "occluder_class",
|
120 |
+
"percentage_occlusion", "mask", "pos"])
|
121 |
+
|
122 |
+
for idx in range(len(dataset)):
|
123 |
+
image, label = dataset[idx]
|
124 |
+
category = dataset.classes[label]
|
125 |
+
|
126 |
+
in_dir = os.path.join(data_dir, category)
|
127 |
+
occluded_dir = os.path.join(occluded_data_dir, category)
|
128 |
+
|
129 |
+
os.makedirs(occluded_dir, exist_ok=True)
|
130 |
+
|
131 |
+
img_name = dataset.imgs[idx][0].split('/')[-1].split('.')[0]
|
132 |
+
|
133 |
+
occluder_tensor, occluder_class = get_random_texture(dataset_occluder)
|
134 |
+
occluded_image, mask_path, pos = occlude_image(image, occluder_tensor, 0.3, occluded_dir, img_name)
|
135 |
+
|
136 |
+
mask_array = np.load(mask_path)
|
137 |
+
actual_percentage_occlusion = np.count_nonzero(mask_array) / (image.shape[1] * image.shape[2])
|
138 |
+
|
139 |
+
occluded_image_path = os.path.join(occluded_dir, f"{img_name}_occluded.png")
|
140 |
+
transforms.ToPILImage()(occluded_image).save(occluded_image_path)
|
141 |
+
|
142 |
+
new_row = pd.DataFrame({
|
143 |
+
"image_name": [f"{img_name}_occluded.png"],
|
144 |
+
"class_name": [category],
|
145 |
+
"occluder_class": [occluder_class],
|
146 |
+
"percentage_occlusion": [actual_percentage_occlusion],
|
147 |
+
"mask": [mask_path],
|
148 |
+
"pos": [pos]
|
149 |
+
})
|
150 |
+
|
151 |
+
occlusion_info = pd.concat([occlusion_info, new_row], ignore_index=True)
|
152 |
+
|
153 |
+
occlusion_info.to_csv(os.path.join(occluded_data_dir, "occlusion_info.csv"), index=False)
|
154 |
+
|
155 |
+
if __name__ == "__main__":
|
156 |
+
main()
|
occlusion_info.csv
ADDED
The diff for this file is too large to render.
See raw diff
|
|