-
Notifications
You must be signed in to change notification settings - Fork 23
/
augmentation.py
252 lines (186 loc) · 8.55 KB
/
augmentation.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
import cv2
import torch
import torchvision
import numpy as np
def color_jitter(image1, image2, brightness, contrast, saturation, hue):
assert image1.shape == image2.shape
cj_module = torchvision.transforms.ColorJitter(brightness, contrast, saturation, hue)
images = np.concatenate([image1, image2], axis=0)
images_t = torch.from_numpy(images.transpose([2, 0, 1]).copy())
images_t = cj_module.forward(images_t / 255.0) * 255.0
images = images_t.numpy().astype(np.uint8).transpose(1, 2, 0)
image1, image2 = images[:image1.shape[0]], images[image1.shape[0]:]
return image1, image2
def flip_point_cloud(pc, image_h, image_w, f, cx, cy, flip_mode):
assert flip_mode in ['lr', 'ud']
pc_x, pc_y, depth = pc[..., 0], pc[..., 1], pc[..., 2]
image_x = cx + (f / depth) * pc_x
image_y = cy + (f / depth) * pc_y
if flip_mode == 'lr':
image_x = image_w - 1 - image_x
else:
image_y = image_h - 1 - image_y
pc_x = (image_x - cx) * depth / f
pc_y = (image_y - cy) * depth / f
pc = np.concatenate([pc_x[:, None], pc_y[:, None], depth[:, None]], axis=-1)
return pc
def flip_scene_flow(pc1, flow_3d, image_h, image_w, f, cx, cy, flip_mode):
new_pc1 = flip_point_cloud(pc1, image_h, image_w, f, cx, cy, flip_mode)
new_pc1_warp = flip_point_cloud(pc1 + flow_3d[:, :3], image_h, image_w, f, cx, cy, flip_mode)
return np.concatenate([new_pc1_warp - new_pc1, flow_3d[:, 3:]], axis=-1)
def flip_image(image, flip_mode):
if flip_mode == 'lr':
return np.fliplr(image).copy()
else:
return np.flipud(image).copy()
def flip_optical_flow(flow, flip_mode):
assert flip_mode in ['lr', 'ud']
if flip_mode == 'lr':
flow = np.fliplr(flow).copy()
flow[:, :, 0] *= -1
else:
flow = np.flipud(flow).copy()
flow[:, :, 1] *= -1
return flow
def random_flip(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, flip_mode):
assert flow_3d.shape[1] <= 4
assert flip_mode in ['lr', 'ud']
image_h, image_w = image1.shape[:2]
if np.random.rand() < 0.5: # do nothing
return image1, image2, pc1, pc2, flow_2d, flow_3d
# flip images
new_image1 = flip_image(image1, flip_mode)
new_image2 = flip_image(image2, flip_mode)
# flip point clouds
new_pc1 = flip_point_cloud(pc1, image_h, image_w, f, cx, cy, flip_mode)
new_pc2 = flip_point_cloud(pc2, image_h, image_w, f, cx, cy, flip_mode)
# flip optical flow and scene flow
new_flow_2d = flip_optical_flow(flow_2d, flip_mode)
new_flow_3d = flip_scene_flow(pc1, flow_3d, image_h, image_w, f, cx, cy, flip_mode)
return new_image1, new_image2, new_pc1, new_pc2, new_flow_2d, new_flow_3d
def crop_image_with_pc(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, crop_window, drop_pc=False):
assert len(crop_window) == 4 # [x1, y1, x2, y2]
x1, y1, x2, y2 = crop_window
image_h, image_w = image1.shape[:2]
# project points to image plane
cx = (image_w - 1) / 2 if cx is None else cx
cy = (image_h - 1) / 2 if cy is None else cy
xyz1_x, xyz1_y, xyz1_z = pc1[..., 0], pc1[..., 1], pc1[..., 2]
xyz2_x, xyz2_y, xyz2_z = pc2[..., 0], pc2[..., 1], pc2[..., 2]
xy1_x = cx + (f / xyz1_z) * xyz1_x
xy1_y = cy + (f / xyz1_z) * xyz1_y
xy2_x = cx + (f / xyz2_z) * xyz2_x
xy2_y = cy + (f / xyz2_z) * xyz2_y
# crop images
image1 = image1[y1:y2, x1:x2].copy()
image2 = image2[y1:y2, x1:x2].copy()
flow_2d = flow_2d[y1:y2, x1:x2].copy()
# crop points
if drop_pc:
crop_mask1 = np.where(np.logical_and(
np.logical_and(xy1_x > x1, xy1_x < x2),
np.logical_and(xy1_y > y1, xy1_y < y2)
))[0]
crop_mask2 = np.where(np.logical_and(
np.logical_and(xy2_x > x1, xy2_x < x2),
np.logical_and(xy2_y > y1, xy2_y < y2)
))[0]
pc1, pc2, flow_3d = pc1[crop_mask1], pc2[crop_mask2], flow_3d[crop_mask1]
if pc1.shape[0] == 0 or pc2.shape[0] == 0:
raise AssertionError
# adjust camera params
cx = cx - x1
cy = cy - y1
return image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy
def random_crop(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, crop_size, drop_pc):
assert flow_3d.shape[1] <= 4
assert len(crop_size) == 2
crop_w, crop_h = crop_size
image_h, image_w = image1.shape[:2]
assert crop_w <= image_w and crop_h <= image_h
# top left of the cropping window
x1 = np.random.randint(low=0, high=image_w - crop_w + 1)
y1 = np.random.randint(low=0, high=image_h - crop_h + 1)
crop_window = [x1, y1, x1 + crop_w, y1 + crop_h]
return crop_image_with_pc(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, crop_window, drop_pc)
def resize_sparse_flow_map(flow, target_w, target_h):
curr_h, curr_w = flow.shape[:2]
coords = np.meshgrid(np.arange(curr_w), np.arange(curr_h))
coords = np.stack(coords, axis=-1).astype(np.float32)
mask = flow[..., -1] > 0
coords0, flow0 = coords[mask], flow[mask][:, :2]
scale_ratio_w = (target_w - 1) / (curr_w - 1)
scale_ratio_h = (target_h - 1) / (curr_h - 1)
coords1 = coords0 * [scale_ratio_w, scale_ratio_h]
flow1 = flow0 * [scale_ratio_w, scale_ratio_h]
xx = np.round(coords1[:, 0]).astype(np.int32)
yy = np.round(coords1[:, 1]).astype(np.int32)
valid = (xx >= 0) & (xx < target_w) & (yy >= 0) & (yy < target_h)
xx, yy, flow1 = xx[valid], yy[valid], flow1[valid]
flow_resized = np.zeros([target_h, target_w, 3], dtype=np.float32)
flow_resized[yy, xx, :2] = flow1
flow_resized[yy, xx, 2:] = 1.0
return flow_resized
def random_scale(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, scale_range):
assert len(scale_range) == 2
assert 1 <= scale_range[0] < scale_range[1]
if np.random.rand() < 0.5:
return image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy
scale_ratio = np.random.uniform(scale_range[0], scale_range[1])
image_h, image_w = image1.shape[:2]
crop_h, crop_w = int(image_h / scale_ratio), int(image_w / scale_ratio)
# top left of the cropping window
x1 = np.random.randint(low=0, high=image_w - crop_w + 1)
y1 = np.random.randint(low=0, high=image_h - crop_h + 1)
crop_window = [x1, y1, x1 + crop_w, y1 + crop_h]
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy = crop_image_with_pc(
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, crop_window
)
# resize images and optical flow
image1 = cv2.resize(image1, (image_w, image_h), interpolation=cv2.INTER_LINEAR)
image2 = cv2.resize(image2, (image_w, image_h), interpolation=cv2.INTER_LINEAR)
flow_2d = resize_sparse_flow_map(flow_2d, image_w, image_h)
# resize points and scene flow
scale_ratio_w = (image_w - 1) / (crop_w - 1)
scale_ratio_h = (image_h - 1) / (crop_h - 1)
pc1[:, 0] *= scale_ratio_w
pc1[:, 1] *= scale_ratio_h
pc2[:, 0] *= scale_ratio_w
pc2[:, 1] *= scale_ratio_h
flow_3d[:, 0] *= scale_ratio_w
flow_3d[:, 1] *= scale_ratio_h
# adjust camera params
cx *= scale_ratio_w
cy *= scale_ratio_h
return image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy
def joint_augmentation(image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, cfgs):
if not cfgs.enabled:
return image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy
if cfgs.color_jitter.enabled:
image1, image2 = color_jitter(
image1, image2,
brightness=cfgs.color_jitter.brightness,
contrast=cfgs.color_jitter.contrast,
saturation=cfgs.color_jitter.saturation,
hue=cfgs.color_jitter.hue,
)
if cfgs.random_horizontal_flip.enabled:
image1, image2, pc1, pc2, flow_2d, flow_3d = random_flip(
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, flip_mode='lr'
)
if cfgs.random_vertical_flip.enabled:
image1, image2, pc1, pc2, flow_2d, flow_3d = random_flip(
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy, flip_mode='ud'
)
if cfgs.random_crop.enabled:
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy = random_crop(
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy,
crop_size=cfgs.random_crop.crop_size,
drop_pc=cfgs.random_crop.drop_pc
)
if cfgs.random_scale.enabled:
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy = random_scale(
image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy,
scale_range=cfgs.random_scale.scale_range
)
return image1, image2, pc1, pc2, flow_2d, flow_3d, f, cx, cy