File size: 8,792 Bytes
625a17f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 |
import json
import os
from pycocotools.coco import COCO
from pycocotools.mask import encode, decode, frPyObjects
import numpy as np
from tqdm import tqdm
from skimage.measure import label, regionprops
from skimage.draw import line
from scipy.ndimage import gaussian_filter
import random
def calculate_iou(box1, box2):
xA = max(box1[1], box2[1])
yA = max(box1[0], box2[0])
xB = min(box1[3], box2[3])
yB = min(box1[2], box2[2])
interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
box1Area = (box1[2] - box1[0] + 1) * (box1[3] - box1[1] + 1)
box2Area = (box2[2] - box2[0] + 1) * (box2[3] - box2[1] + 1)
iou = interArea / float(box1Area + box2Area - interArea)
return iou
def generate_visual_prompt(mask):
# point
label_mask = label(mask)
props_ = regionprops(label_mask)
props = []
for prop in props_:
if prop.area > 5:
props.append(prop)
point_visual_prompt_mask = np.zeros_like(mask, dtype=np.uint8)
for prop in props:
# Randomly choose a point within the region
min_row, min_col, max_row, max_col = prop.bbox
centroid = prop.centroid
for _ in range(1000):
radius = min(prop.bbox[2] - prop.bbox[0], prop.bbox[3] - prop.bbox[1]) * 0.5
angle = random.uniform(0, 2 * np.pi)
offset = (random.uniform(0, radius) * np.cos(angle), random.uniform(0, radius) * np.sin(angle))
point = (int(centroid[0] + offset[0]), int(centroid[1] + offset[1]))
point = (np.clip(point[0], min_row, max_row - 1), np.clip(point[1], min_col, max_col - 1))
if mask[point[0], point[1]] > 0:
point_visual_prompt_mask[point[0], point[1]] = 1
break
# mask
mask_visual_prompt_mask = gaussian_filter(mask.astype(float), sigma=2)
mask_visual_prompt_mask = (mask_visual_prompt_mask > mask_visual_prompt_mask.mean()).astype(np.uint8)
# box
box_visual_prompt_mask = np.zeros_like(mask, dtype=np.uint8)
for prop in props:
min_row, min_col, max_row, max_col = prop.bbox
scale_factor = random.uniform(0.9, 1.1)
height = max_row - min_row
width = max_col - min_col
delta_height = height * (scale_factor - 1)
delta_width = width * (scale_factor - 1)
min_row = max(0, int(min_row - delta_height / 2))
min_col = max(0, int(min_col - delta_width / 2))
max_row = min(mask.shape[0], int(max_row + delta_height / 2))
max_col = min(mask.shape[1], int(max_col + delta_width / 2))
box_visual_prompt_mask[min_row:max_row, min_col:max_col] = 1
# scribble
scribble_visual_prompt_mask = np.zeros_like(mask, dtype=np.uint8)
for prop in props:
min_row, min_col, max_row, max_col = prop.bbox
center_row, center_col = prop.centroid
new_height = (max_row - min_row) * random.uniform(0.5, 1.2)
new_width = (max_col - min_col) * random.uniform(0.5, 1.2)
new_min_row = int(center_row - new_height / 2)
new_min_col = int(center_col - new_width / 2)
new_max_row = int(center_row + new_height / 2)
new_max_col = int(center_col + new_width / 2)
new_min_row, new_min_col = max(new_min_row, 0), max(new_min_col, 0)
new_max_row, new_max_col = min(new_max_row, mask.shape[0]), min(new_max_col, mask.shape[1])
new_box = (new_min_row, new_min_col, new_max_row, new_max_col)
original_box = (min_row, min_col, max_row, max_col)
flag = True
for _ in range(1000):
if calculate_iou(new_box, original_box) < 0.5:
new_height = (max_row - min_row) * random.uniform(0.5, 1.2)
new_width = (max_col - min_col) * random.uniform(0.5, 1.2)
new_min_row = int(center_row - new_height / 2)
new_min_col = int(center_col - new_width / 2)
new_max_row = int(center_row + new_height / 2)
new_max_col = int(center_col + new_width / 2)
new_min_row, new_min_col = max(new_min_row, 0), max(new_min_col, 0)
new_max_row, new_max_col = min(new_max_row, mask.shape[0]), min(new_max_col, mask.shape[1])
new_box = (new_min_row, new_min_col, new_max_row, new_max_col)
else:
flag = False
break
if flag:
continue
corners = [(new_min_row, new_min_col), (new_min_row, new_max_col),
(new_max_row, new_min_col), (new_max_row, new_max_col)]
start_point = random.choice(corners)
corners.remove(start_point)
if start_point in [(new_min_row, new_min_col), (new_max_row, new_max_col)]:
end_point = (new_max_row if start_point[0] == new_min_row else new_min_row,
new_max_col if start_point[1] == new_min_col else new_min_col)
else:
end_point = (new_max_row if start_point[0] == new_min_row else new_min_row,
new_min_col if start_point[1] == new_max_col else new_max_col)
rr, cc = line(start_point[0], start_point[1], end_point[0], end_point[1])
rr = np.array(rr, dtype=np.float32)
cc = np.array(cc, dtype=np.float32)
amplitude = random.uniform(10, 20)
frequency = random.uniform(0.2, 1)
phase_shift = random.uniform(0, 2 * np.pi)
sine_wave = amplitude * np.sin(2 * np.pi * frequency * np.linspace(0, 1, len(rr)) + phase_shift)
rr += sine_wave
rr = np.clip(rr, 0, mask.shape[0] - 1).astype(np.int32)
cc = np.clip(cc, 0, mask.shape[1] - 1).astype(np.int32)
scribble_visual_prompt_mask[rr, cc] = 1
return point_visual_prompt_mask, mask_visual_prompt_mask, box_visual_prompt_mask, scribble_visual_prompt_mask
if __name__ == '__main__':
root_path = 'datasets/coco'
splits = ['train', 'val']
for split in splits:
print(f'Processing {split}...')
coco_path = os.path.join(root_path, f'annotation/instances_{split}2017.json')
save_path = os.path.join(root_path, f'coco_interactive_{split}_psalm.json')
coco = COCO(coco_path)
coco_interactivate = []
new_img_id = 0
for img_id in tqdm(coco.imgs):
img_info = coco.imgs[img_id]
ann_ids = coco.getAnnIds(imgIds=img_id)
anns = coco.loadAnns(ann_ids)
if len(anns) == 0:
print('no annotation')
continue
for ann in anns:
if isinstance(ann['segmentation'], list):
rle = frPyObjects(ann['segmentation'], img_info['height'], img_info['width'])
mask = coco.annToMask(ann)
elif isinstance(ann['segmentation'], dict):
mask = coco.annToMask(ann)
else:
raise ValueError("Unknown segmentation format")
point_visual_prompt_mask, mask_visual_prompt_mask, box_visual_prompt_mask, scribble_visual_prompt_mask = generate_visual_prompt(
mask)
point_rle = encode(np.asfortranarray(point_visual_prompt_mask))
mask_rle = encode(np.asfortranarray(mask_visual_prompt_mask))
box_rle = encode(np.asfortranarray(box_visual_prompt_mask))
scribble_rle = encode(np.asfortranarray(scribble_visual_prompt_mask))
ann['point_visual_prompt_mask'] = {
'counts': point_rle['counts'].decode('ascii'),
'size': point_rle['size']
}
ann['mask_visual_prompt_mask'] = {
'counts': mask_rle['counts'].decode('ascii'),
'size': mask_rle['size']
}
ann['box_visual_prompt_mask'] = {
'counts': box_rle['counts'].decode('ascii'),
'size': box_rle['size']
}
ann['scribble_visual_prompt_mask'] = {
'counts': scribble_rle['counts'].decode('ascii'),
'size': scribble_rle['size']
}
coco_interactivate.append({
'image': img_info['file_name'],
'image_info': img_info,
'new_img_id': new_img_id,
'anns': anns
})
new_img_id += 1
with open(save_path, 'w') as f:
json.dump(coco_interactivate, f, indent=2)
print('dataset save in {}, max new_img_id: {}'.format(save_path, new_img_id))
|