-
Notifications
You must be signed in to change notification settings - Fork 31
/
augment.py
427 lines (358 loc) · 13.7 KB
/
augment.py
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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
import random
import numpy as np
from scipy import ndimage
import chainercv.transforms as transforms
from chainercv.links.model.ssd.transforms import random_distort
import PIL
from PIL import ImageChops, ImageOps, ImageFilter, ImageEnhance
def rotate_point(point_yx, angle, center_yx):
offset_y, offset_x = center_yx
shift = point_yx - center_yx
shift_y, shift_x = shift[:, 0], shift[:, 1]
cos_rad = np.cos(np.deg2rad(angle))
sin_rad = np.sin(np.deg2rad(angle))
qx = offset_x + cos_rad * shift_x + sin_rad * shift_y
qy = offset_y - sin_rad * shift_x + cos_rad * shift_y
return np.array([qy, qx]).transpose()
def rotate_image(image, angle):
rot = ndimage.rotate(image, angle, axes=(2, 1), reshape=False)
# disable image collapse
rot = np.clip(rot, 0, 255)
return rot
def random_rotate(image, keypoints, bbox):
angle = np.random.uniform(-40, 40)
param = {}
param['angle'] = angle
new_keypoints = []
center_yx = np.array(image.shape[1:]) / 2
for points in keypoints:
rot_points = rotate_point(np.array(points),
angle,
center_yx)
new_keypoints.append(rot_points)
new_bbox = []
for x, y, w, h in bbox:
points = np.array(
[
[y, x],
[y, x + w],
[y + h, x],
[y + h, x + w]
]
)
rot_points = rotate_point(
points,
angle,
center_yx
)
xmax = np.max(rot_points[:, 1])
ymax = np.max(rot_points[:, 0])
xmin = np.min(rot_points[:, 1])
ymin = np.min(rot_points[:, 0])
# x,y,w,h
new_bbox.append([xmin, ymin, xmax - xmin, ymax - ymin])
image = rotate_image(image, angle)
return image, new_keypoints, new_bbox, param
def spot_light(pil_img):
w, h = pil_img.size
effect_img = np.zeros((h, w, 3))
scale_w = random.choice([5, 6, 7, 8, 9])
scale_h = random.choice([5, 6, 7, 8, 9])
x = random.choice(range(w // scale_w, w - w // scale_w))
y = random.choice(range(h // scale_h, h - h // scale_h))
light = random.choice(range(128, 220))
effect_img[y - h // scale_h:y + h // scale_h, x - w // scale_w:x + w // scale_w] = light
effect_img = PIL.Image.fromarray(effect_img.astype(np.uint8))
return ImageChops.add(pil_img, effect_img)
def blend_alpha(pil_img, direction='left'):
w, h = pil_img.size
effect_img = np.zeros((h, w, 3))
if direction == 'right':
for x in range(w):
effect_img[:, x] = x * 255 / w
elif direction == 'left':
for x in range(w):
effect_img[:, x] = (w - x) * 255 / w
elif direction == 'up':
for y in range(h):
effect_img[y, :] = (h - y) * 255 / h
elif direction == 'down':
for y in range(h):
effect_img[y, :] = y * 255 / h
else:
raise Exception("invalid argument direction is 'right','left','up','down' actual {}".format(direction))
effect_img = PIL.Image.fromarray(effect_img.astype(np.uint8))
return PIL.Image.blend(pil_img, effect_img, 0.5)
def chop_image(pil_img, direction='left', op='add'):
w, h = pil_img.size
effect_img = np.zeros((h, w, 3))
if direction == 'right':
for x in range(w):
effect_img[:, x] = x * 255 / w
elif direction == 'left':
for x in range(w):
effect_img[:, x] = (w - x) * 255 / w
elif direction == 'up':
for y in range(h):
effect_img[y, :] = (h - y) * 255 / h
elif direction == 'down':
for y in range(h):
effect_img[y, :] = y * 255 / h
else:
raise Exception("invalid argument direction. It should be 'right','left','up','down' actual {}".format(direction))
effect_img = PIL.Image.fromarray(effect_img.astype(np.uint8))
if op == 'add':
operation = ImageChops.add
elif op == 'subtract':
operation = ImageChops.subtract
elif op == 'multiply':
operation = ImageChops.multiply
elif op == 'screen':
operation = ImageChops.screen
elif op == 'lighter':
operation = ImageChops.lighter
elif op == 'darker':
operation = ImageChops.darker
else:
ops = ['add', 'subtract', 'multiply', 'screen', 'lighter', 'darker']
raise Exception("invalid argument op. {} actual {}".format(ops, direction))
return operation(pil_img, effect_img)
def filter_image(pil_img):
method = random.choice(['gaussian', 'blur', 'sharpen'])
if method == 'gaussian':
return pil_img.filter(ImageFilter.GaussianBlur(random.choice([0.5, 1.0, 1.5])))
if method == 'blur':
return pil_img.filter(ImageFilter.BLUR)
if method == 'sharpen':
return pil_img.filter(ImageFilter.SHARPEN)
def random_process_by_PIL(image):
# convert CHW -> HWC -> PIL.Image
pil_img = PIL.Image.fromarray(image.transpose(1, 2, 0).astype(np.uint8))
method = np.random.choice(
['equalize', 'spot_light', 'chop', 'blend'],
p=[0.15, 0.15, 0.35, 0.35]
)
param = {'pil': method, 'filter': False}
if method == 'equalize':
pil_img = ImageOps.equalize(pil_img)
if method == 'spot_light':
pil_img = spot_light(pil_img)
if method == 'chop':
direction = random.choice(['left', 'right', 'up', 'down'])
op = random.choice(['add', 'subtract', 'multiply', 'screen', 'lighter', 'darker'])
pil_img = chop_image(pil_img, direction, op)
if method == 'blend':
direction = random.choice(['left', 'right', 'up', 'down'])
pil_img = blend_alpha(pil_img, direction)
if np.random.choice([True, False], p=[0.1, 0.9]):
pil_img = filter_image(pil_img)
param['filter'] = True
# back to CHW
image = np.asarray(pil_img).transpose(2, 0, 1).astype(np.float32)
return image, param
def augment_image(image, dataset_type):
"""color augmentation"""
param = {}
if dataset_type == 'mpii':
method = np.random.choice(
['random_distort', 'pil'],
p=[1.0, 0.0],
)
elif dataset_type == 'coco':
method = np.random.choice(
['random_distort', 'pil'],
p=[0.5, 0.5],
)
if method == 'random_distort':
param['method'] = method
# color augmentation provided by ChainerCV
ret = random_distort(image, contrast_low=0.3, contrast_high=2)
return ret, param
if method == 'pil':
ret, param = random_process_by_PIL(image)
param['method'] = method
return ret, param
def random_flip(image, keypoints, bbox, is_labeled, is_visible, flip_indices):
"""
random x_flip
Note that if image is flipped, `flip_indices` translate elements.
e.g. left_shoulder -> right_shoulder.
"""
_, H, W = image.shape
image, param = transforms.random_flip(image, x_random=True, return_param=True)
if param['x_flip']:
keypoints = [
transforms.flip_point(points, (H, W), x_flip=True)[flip_indices]
for points in keypoints
]
is_labeled = [label[flip_indices] for label in is_labeled]
is_visible = [vis[flip_indices] for vis in is_visible]
new_bbox = []
for x, y, w, h in bbox:
[[y, x]] = transforms.flip_point(np.array([[y, x + w]]), (H, W), x_flip=True)
new_bbox.append([x, y, w, h])
bbox = new_bbox
return image, keypoints, bbox, is_labeled, is_visible, param
def scale_fit_short(image, keypoints, bbox, length):
_, H, W = image.shape
min_hw = min(H, W)
scale = length / min_hw
new_image = transforms.scale(image, size=length, fit_short=True)
new_keypoints = [scale * k for k in keypoints]
new_bbox = [scale * np.asarray(b) for b in bbox]
return new_image, new_keypoints, new_bbox
def intersection(bbox0, bbox1):
x0, y0, w0, h0 = bbox0
x1, y1, w1, h1 = bbox1
def relu(x): return max(0, x)
w = relu(min(x0 + w0, x1 + w1) - max(x0, x1))
h = relu(min(y0 + h0, y1 + h1) - max(y0, y1))
return w * h
def translate_bbox(bbox, size, y_offset, x_offset):
cropped_H, cropped_W = size
new_bbox = []
for x, y, w, h in bbox:
x_shift = x + x_offset
y_shift = y + y_offset
is_intersect = intersection([0, 0, cropped_W, cropped_H], [x_shift, y_shift, w, h])
if is_intersect:
xmin = max(0, x_shift)
ymin = max(0, y_shift)
xmax = min(cropped_W, x_shift + w)
ymax = min(cropped_H, y_shift + h)
wnew = xmax - xmin
hnew = ymax - ymin
new_bbox.append([xmin, ymin, wnew, hnew])
else:
new_bbox.append([x_shift, y_shift, w, h])
return new_bbox
def crop(img, y_slice, x_slice, copy=False):
ret = img.copy() if copy else img
return ret[:, y_slice, x_slice]
def crop_all_humans(image, keypoints, bbox, is_labeled):
_, H, W = image.shape
aspect = W / H
param = {}
if len(keypoints) == 0:
param['do_nothing'] = True
return image, keypoints, bbox, param
kymax = max([np.max(ks[l, 0]) for l, ks in zip(is_labeled, keypoints)])
kxmax = max([np.max(ks[l, 1]) for l, ks in zip(is_labeled, keypoints)])
kymin = min([np.min(ks[l, 0]) for l, ks in zip(is_labeled, keypoints)])
kxmin = min([np.min(ks[l, 1]) for l, ks in zip(is_labeled, keypoints)])
bxmax = max([b[0] + b[2] for b in bbox])
bymax = max([b[1] + b[3] for b in bbox])
bxmin = min([b[0] for b in bbox])
bymin = min([b[1] for b in bbox])
ymax = max(kymax, bymax)
xmax = max(kxmax, bxmax)
ymin = min(kymin, bymin)
xmin = min(kxmin, bxmin)
if (xmax + xmin) / 2 < W / 2:
x_start = random.randint(0, max(0, int(xmin)))
y_start = random.randint(0, max(0, int(ymin)))
y_end = random.randint(min(H, int(ymax)), H)
ylen = y_end - y_start
xlen = aspect * ylen
x_end = min(W, int(x_start + xlen))
x_slice = slice(x_start, x_end, None)
y_slice = slice(y_start, y_end, None)
else:
x_end = random.randint(min(int(xmax), W), W)
y_end = random.randint(min(int(ymax), H), H)
y_start = random.randint(0, max(0, int(ymin)))
ylen = y_end - y_start
xlen = aspect * ylen
x_start = max(0, int(x_end - xlen))
x_slice = slice(x_start, x_end, None)
y_slice = slice(y_start, y_end, None)
cropped = crop(image, y_slice=y_slice, x_slice=x_slice, copy=True)
_, cropped_H, cropped_W = cropped.shape
param['x_slice'] = x_slice
param['y_slice'] = y_slice
if cropped_H <= 50 or cropped_W <= 50:
"""
This case, for example, cropped_H=0 will cause an error when try to resize image
or resize small image to insize will cause low resolution human image.
To avoid situations, we will stop crop image.
"""
param['do_nothing'] = True
return image, keypoints, bbox, param
image = cropped
keypoints = [
transforms.translate_point(
points, x_offset=-x_slice.start, y_offset=-y_slice.start)
for points in keypoints
]
bbox = translate_bbox(
bbox,
size=(cropped_H, cropped_W),
x_offset=-x_slice.start,
y_offset=-y_slice.start,
)
return image, keypoints, bbox, param
def random_sized_crop(image, keypoints, bbox):
image, param = transforms.random_sized_crop(
image,
scale_ratio_range=(0.5, 5),
aspect_ratio_range=(0.75, 1.3333333333333333),
return_param=True
)
keypoints = [
transforms.translate_point(points,
x_offset=-param['x_slice'].start,
y_offset=-param['y_slice'].start
)
for points in keypoints
]
_, cropped_H, cropped_W = image.shape
bbox = translate_bbox(
bbox,
size=(cropped_H, cropped_W),
x_offset=-param['x_slice'].start,
y_offset=-param['y_slice'].start,
)
return image, keypoints, bbox, {random_sized_crop.__name__: param}
def resize(image, keypoints, bbox, size):
_, H, W = image.shape
new_h, new_w = size
image = transforms.resize(image, (new_h, new_w))
keypoints = [
transforms.resize_point(points, (H, W), (new_h, new_w))
for points in keypoints
]
new_bbox = []
for x, y, bw, bh in bbox:
[[y, x]] = transforms.resize_point(np.array([[y, x]]), (H, W), (new_h, new_w))
bw *= new_w / W
bh *= new_h / H
new_bbox.append([x, y, bw, bh])
return image, keypoints, new_bbox
def random_resize(image, keypoints, bbox):
# Random resize
_, H, W = image.shape
scalew, scaleh = np.random.uniform(0.7, 1.5, 2)
resizeW, resizeH = int(W * scalew), int(H * scaleh)
image, keypoints, bbox = resize(image, keypoints, bbox, (resizeH, resizeW))
return image, keypoints, bbox, {'H': resizeH, 'W': resizeW}
def random_crop(image, keypoints, bbox, is_labeled, dataset_type):
if dataset_type == 'mpii':
crop_target = np.random.choice(
['random_sized_crop', 'crop_all_humans'],
p=[0.1, 0.9],
)
if dataset_type == 'coco':
crop_target = np.random.choice(
['random_sized_crop', 'crop_all_humans'],
p=[0.5, 0.5],
)
param = {'crop_target': crop_target}
if crop_target == 'random_sized_crop':
image, keypoints, bbox, p = random_resize(image, keypoints, bbox)
param.update(p)
image, keypoints, bbox, p = random_sized_crop(image, keypoints, bbox)
param.update(p)
elif crop_target == 'crop_all_humans':
image, keypoints, bbox, p = crop_all_humans(image, keypoints, bbox, is_labeled)
param.update(p)
return image, keypoints, bbox, param