-
Notifications
You must be signed in to change notification settings - Fork 19
/
Copy pathnet_utils.py
306 lines (259 loc) · 12.7 KB
/
net_utils.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
"""
Created on April, 2018
@author: Siyuan Huang
pytorch network utils
"""
import torch
import torch.nn as nn
import numpy as np
import copy
PI = 3.141592653589793
def conv(in_plane, out_plane, kernel_size, stride):
"""
conv2d layer with same padding
:param in_plane:
:param out_plane:
:param kernel_size:
:return: conv with same padding
"""
return nn.Sequential(
nn.Conv2d(in_plane, out_plane, kernel_size=kernel_size, padding=(kernel_size - 1) // stride, stride=stride),
)
def basis_from_ori(ori):
"""
:param ori: torch tensor
the orientation angle
:return: basis: 3x3 tensor
the basis in 3D coordinates
"""
n = ori.size(0)
ori = ori / 180.0 * PI
basis = torch.zeros((n, 3, 3)).cuda()
basis[:, 0, 0] = torch.cos(ori)
basis[:, 0, 1] = - torch.sin(ori)
basis[:, 1, 0] = torch.sin(ori)
basis[:, 1, 1] = torch.cos(ori)
basis[:, 2, 2] = 1
return basis
def get_corners_of_bb3d(basis, coeffs, centroid):
"""
:param basis: b x 3 x 3 tensor
:param coeffs: b x 3 tensor
:param centroid: b x 3 tensor
:return: corners b x 8 x 3 tensor
"""
n = basis.size(0)
corners = torch.zeros((n, 8, 3)).cuda()
coeffs = coeffs.view(n, 3, 1).expand(-1, -1, 3)
centroid = centroid.view(n, 1, 3).expand(-1, 8, -1)
corners[:, 0, :] = -basis[:, 0, :] * coeffs[:, 0, :] + basis[:, 1, :] * coeffs[:, 1, :] + basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 1, :] = basis[:, 0, :] * coeffs[:, 0, :] + basis[:, 1, :] * coeffs[:, 1, :] + basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 2, :] = basis[:, 0, :] * coeffs[:, 0, :] + -basis[:, 1, :] * coeffs[:, 1, :] + basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 3, :] = -basis[:, 0, :] * coeffs[:, 0, :] + -basis[:, 1, :] * coeffs[:, 1, :] + basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 4, :] = -basis[:, 0, :] * coeffs[:, 0, :] + basis[:, 1, :] * coeffs[:, 1, :] + -basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 5, :] = basis[:, 0, :] * coeffs[:, 0, :] + basis[:, 1, :] * coeffs[:, 1, :] + -basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 6, :] = basis[:, 0, :] * coeffs[:, 0, :] + -basis[:, 1, :] * coeffs[:, 1, :] + -basis[:, 2, :] * coeffs[:, 2, :]
corners[:, 7, :] = -basis[:, 0, :] * coeffs[:, 0, :] + -basis[:, 1, :] * coeffs[:, 1, :] + -basis[:, 2, :] * coeffs[:, 2, :]
corners = corners + centroid
return corners
def to_dict_tensor(dicts, if_cuda):
dicts_new = copy.copy(dicts)
for key, value in dicts_new.items():
value_new = torch.from_numpy(np.array(value))
if value_new.type() == 'torch.DoubleTensor':
value_new = value_new.float()
if if_cuda:
value_new = value_new.cuda()
dicts_new[key] = value_new
return dicts_new
def physical_violation(bdb_layout, bdb_3d):
"""
compute the loss of physical violation
:param bdb_layout: 1 x 8 x 3 tensor
:param bdb_3d: b x 8 x 3 tensor
:return:
"""
b = bdb_3d.size(0)
layout_max = torch.max(bdb_layout, dim=1)[0].expand(b, -1) # bx3
layout_min = torch.min(bdb_layout, dim=1)[0].expand(b, -1) # bx3
bdb_max = torch.max(bdb_3d, dim=1)[0] # b x 3
bdb_min = torch.min(bdb_3d, dim=1)[0] # b x 3
violation = torch.nn.functional.relu(bdb_max - layout_max) + torch.nn.functional.relu(layout_min - bdb_min) # b x 3
return violation, torch.zeros(b, 3).cuda()
def get_rotation_matix_result(bins_tensor, yaw_cls_gt, yaw_reg_result, roll_cls_gt, roll_reg_result):
"""
:param bins_tensor:
:param yaw_cls_gt:
:param yaw_reg_result:
:param roll_cls_gt:
:param roll_reg_result:
:return: r_ex 1 x 3 x 3 tensor
"""
yaw_result = torch.gather(yaw_reg_result, 1, yaw_cls_gt.view(yaw_cls_gt.size(0), 1).expand(yaw_cls_gt.size(0), 1)).squeeze(1)
roll_result = torch.gather(roll_reg_result, 1, roll_cls_gt.view(roll_cls_gt.size(0), 1).expand(roll_cls_gt.size(0), 1)).squeeze(1)
yaw = num_from_bins(bins_tensor['yaw_bin'], yaw_cls_gt, yaw_result)
roll = num_from_bins(bins_tensor['roll_bin'], roll_cls_gt, roll_result)
r_ex = get_rotation_matrix_from_yaw_roll(-yaw / 180.0 * PI, -roll / 180.0 * PI)
return r_ex
def get_rotation_matrix_gt(bins_tensor, yaw_cls_gt, yaw_reg_gt, roll_cls_gt, roll_reg_gt):
yaw = num_from_bins(bins_tensor['yaw_bin'], yaw_cls_gt, yaw_reg_gt)
roll = num_from_bins(bins_tensor['roll_bin'], roll_cls_gt, roll_reg_gt)
r_ex = get_rotation_matrix_from_yaw_roll(- yaw / 180.0 * PI, - roll / 180.0 * PI)
return r_ex
def get_bdb_3d_result(bins_tensor, ori_cls_gt, ori_reg_result, centroid_cls_gt, centroid_reg_result, size_cls_gt, size_reg_result, P, K, R_ex):
# coeffs
size_reg = torch.gather(size_reg_result, 1, size_cls_gt.view(size_cls_gt.size(0), 1, 1).expand(size_cls_gt.size(0), 1, size_reg_result.size(2))).squeeze(1)
coeffs = (size_reg + 1) * bins_tensor['avg_size'][size_cls_gt, :] # b x 3
# centroid
centroid_reg = torch.gather(centroid_reg_result, 1, centroid_cls_gt.view(centroid_cls_gt.size(0), 1).expand(centroid_cls_gt.size(0), 1)).squeeze(1)
centroid_depth = num_from_bins(bins_tensor['centroid_bin'], centroid_cls_gt, centroid_reg)
centroid = rgbd_to_world(P, centroid_depth, K, R_ex) # b x 3
# basis
ori_reg = torch.gather(ori_reg_result, 1, ori_cls_gt.view(ori_cls_gt.size(0), 1).expand(ori_cls_gt.size(0), 1)).squeeze(1)
ori = num_from_bins(bins_tensor['ori_bin'], ori_cls_gt, ori_reg)
basis = basis_from_ori(ori)
bdb = get_corners_of_bb3d(basis, coeffs, centroid)
return bdb
def get_bdb_3d_gt(bins_tensor, ori_cls_gt, ori_reg_gt, centroid_cls_gt, centroid_reg_gt, size_cls_gt, size_reg_gt, P, K, R_ex):
# coeffs
coeffs = (size_reg_gt + 1) * bins_tensor['avg_size'][size_cls_gt, :] # b x 3
# centroid
centroid_depth = num_from_bins(bins_tensor['centroid_bin'], centroid_cls_gt, centroid_reg_gt)
centroid = rgbd_to_world(P, centroid_depth, K, R_ex) # b x 3
# basis
ori = num_from_bins(bins_tensor['ori_bin'], ori_cls_gt, ori_reg_gt)
basis = basis_from_ori(ori)
bdb = get_corners_of_bb3d(basis, coeffs, centroid)
return bdb
def rgbd_to_world(p, depth, K, R_ex):
"""
Given pixel location and depth, get world coordinates
:param p: b x 2 tensor
:param depth: b tensor
:param k: 1 x 3 x 3 tensor
:param r_ex: 1 x 3 x 3 tensor
:return: p_world_right: b x 3 tensor in right hand coordinate
"""
n = p.size(0)
x_temp = (p[:, 0] + 1 - K[0, 0, 2]) / K[0, 0, 0]
y_temp = (p[:, 1] + 1 - K[0, 1, 2]) / K[0, 1, 1]
z_temp = 1
x = x_temp / torch.sqrt(x_temp**2 + y_temp**2 + z_temp**2) * depth
y = y_temp / torch.sqrt(x_temp**2 + y_temp**2 + z_temp**2) * depth
z = z_temp / torch.sqrt(x_temp**2 + y_temp**2 + z_temp**2) * depth
p_cam = torch.stack((x, y, z), 1).view(n, 3, 1) # n x 3
p_world = torch.bmm(torch.transpose(R_ex, 1, 2).expand(n, -1, -1), p_cam)
p_world_right = torch.stack((p_world[:, 0, 0], p_world[:, 2, 0], -p_world[:, 1, 0]), 1)
return p_world_right
def get_layout_bdb(bins_tensor, ori_cls, ori_reg, centroid_reg, coeffs_reg):
"""
get the eight corners of 3D bounding box
:param bins_tensor:
:param ori_cls: 1 tensor
:param ori_reg: 1 x 3 tensor
:param centroid_reg: 1 x 3 tensor
:param coeffs_reg: 1 x 3 tensor
:return: bdb
"""
ori_reg = torch.gather(ori_reg, 1, ori_cls.view(1, 1).expand(1, 1)).squeeze(1)
ori = num_from_bins(bins_tensor['layout_ori_bin'], ori_cls, ori_reg)
basis = basis_from_ori(ori)
coeffs_reg = (coeffs_reg + 1) * bins_tensor['layout_coeffs_avg']
centroid_reg = centroid_reg * bins_tensor['layout_normalize'] + bins_tensor['layout_centroid_avg']
bdb = get_corners_of_bb3d(basis, coeffs_reg, centroid_reg)
return bdb
def get_bdb_2d_result(bdb3d, r_ex, K):
"""
:param bins_tensor:
:param bdb3d: b x 8 x 3 tensor
:param r_ex:
:param K: 1 x 3 x 3 tensor
:return:
"""
n = bdb3d.size(0)
points_2d = project_3d_points_to_2d(bdb3d, r_ex, K) # b x 8 x 2
x1 = torch.max(torch.min(points_2d[:, :, 0], dim=1)[0], torch.ones(n).cuda()) / (K[0, 0, 2].float())
y1 = torch.max(torch.min(points_2d[:, :, 1], dim=1)[0], torch.ones(n).cuda()) / (K[0, 1, 2].float())
x2 = torch.min(torch.max(points_2d[:, :, 0], dim=1)[0], torch.ones(n).cuda() * 2 * K[0, 0, 2]) / (K[0, 0, 2].float())
y2 = torch.min(torch.max(points_2d[:, :, 1], dim=1)[0], torch.ones(n).cuda() * 2 * K[0, 1, 2]) / (K[0, 1, 2].float())
return torch.stack((x1, y1, x2, y2), 1)
def get_bdb_2d_gt(bins_tensor, bdb3d, yaw_cls_gt, yaw_reg_gt, roll_cls_gt, roll_reg_gt, K):
n = bdb3d.size(0)
yaw = num_from_bins(bins_tensor['yaw_bin'], yaw_cls_gt, yaw_reg_gt)
roll = num_from_bins(bins_tensor['roll_bin'], roll_cls_gt, roll_reg_gt)
r_ex = get_rotation_matrix_from_yaw_roll(- yaw / 180.0 * PI, - roll / 180.0 * PI)
points_2d = project_3d_points_to_2d(bdb3d, r_ex, K) # b x 8 x 2
x1 = torch.max(torch.min(points_2d[:, :, 0], dim=1)[0], torch.ones(n).cuda())
y1 = torch.max(torch.min(points_2d[:, :, 1], dim=1)[0], torch.ones(n).cuda())
x2 = torch.min(torch.max(points_2d[:, :, 0], dim=1)[0], torch.ones(n).cuda() * 2 * K[0, 0, 2])
y2 = torch.min(torch.max(points_2d[:, :, 1], dim=1)[0], torch.ones(n).cuda() * 2 * K[0, 1, 2])
return torch.stack((x1, y1, x2, y2), 1)
def project_3d_points_to_2d(points3d, R_ex, K):
"""
project 3d points to 2d
:param points3d: b x 8 x 3 tensor
:param R_ex: 1 x 3 x 3 tensor
:param K: 1 x 3 x 3 tensor
:return:
"""
n = points3d.size(0)
points3d_left_hand = torch.stack((points3d[:, :, 0], -points3d[:, :, 2], points3d[:, :, 1]), 2) # transform to left hand coordinate
points_cam_ori = torch.transpose(torch.bmm(R_ex.expand(n, -1, -1), torch.transpose(points3d_left_hand, 1, 2)), 1, 2) # b x 8 x 3
points_cam = torch.transpose(torch.stack((points_cam_ori[:, :, 0], points_cam_ori[:, :, 1], torch.abs(points_cam_ori[:, :, 2])), 2), 1, 2) # b x 3 x 8
points_2d_ori = torch.transpose(torch.bmm(K.expand(n, -1, -1), points_cam), 1, 2) # b x 8 x 3
points_2d = torch.stack((points_2d_ori[:, :, 0] / points_2d_ori[:, :, 2], points_2d_ori[:, :, 1] / points_2d_ori[:, :, 2]), 2) # n x 8 x 2
return points_2d
def get_rotation_matrix_from_yaw_roll(yaw, roll):
"""
:param yaw: b x 1 tensor
:param roll: b x 1 tensor
:return: R_ex in left hand coordinates
"""
n = yaw.size(0)
R = torch.zeros((n, 3, 3)).cuda()
R[:, 0, 0] = torch.cos(yaw)
R[:, 0, 1] = - torch.sin(yaw) * torch.cos(roll)
R[:, 0, 2] = torch.sin(roll) * torch.sin(yaw)
R[:, 1, 0] = torch.sin(yaw)
R[:, 1, 1] = torch.cos(roll) * torch.cos(yaw)
R[:, 1, 2] = - torch.cos(yaw) * torch.sin(roll)
R[:, 2, 0] = 0
R[:, 2, 1] = torch.sin(roll)
R[:, 2, 2] = torch.cos(roll)
return R
def num_from_bins(bins, cls, reg):
"""
:param bins: b x 2 tensors
:param cls: b long tensors
:param reg: b tensors
:return: bin_center: b tensors
"""
bin_width = (bins[0][1] - bins[0][0])
bin_center = (bins[cls, 0] + bins[cls, 1]) / 2
return bin_center + reg * bin_width
"""
Functions for testing
"""
def get_bdb_evaluation(bins_tensor, ori_cls_gt, ori_reg_result, centroid_cls_gt, centroid_reg_result, size_cls_gt, size_reg_result, P, K, R_ex):
size_reg = torch.gather(size_reg_result, 1, size_cls_gt.view(size_cls_gt.size(0), 1, 1).expand(size_cls_gt.size(0), 1, size_reg_result.size(2))).squeeze(1)
coeffs = (size_reg + 1) * bins_tensor['avg_size'][size_cls_gt, :] # b x 3
# centroid
centroid_reg = torch.gather(centroid_reg_result, 1, centroid_cls_gt.view(centroid_cls_gt.size(0), 1).expand(centroid_cls_gt.size(0), 1)).squeeze(1)
centroid_depth = num_from_bins(bins_tensor['centroid_bin'], centroid_cls_gt, centroid_reg)
centroid = rgbd_to_world(P, centroid_depth, K, R_ex) # b x 3
# basis
ori_reg = torch.gather(ori_reg_result, 1, ori_cls_gt.view(ori_cls_gt.size(0), 1).expand(ori_cls_gt.size(0), 1)).squeeze(1)
ori = num_from_bins(bins_tensor['ori_bin'], ori_cls_gt, ori_reg)
basis = basis_from_ori(ori)
n = ori_cls_gt.size(0)
bdb_output = [{'basis': basis[i, :, :].cpu().numpy(), 'coeffs': coeffs[i, :].cpu().numpy(), 'centroid': centroid[i, :].cpu().numpy(), 'classid':size_cls_gt[i].cpu().numpy() + 1} for i in range(n)]
return bdb_output
def get_yaw_roll_error(bin, reg_result, cls_result, reg_gt, cls_gt):
result = torch.gather(reg_result, 1, torch.argmax(cls_result, 1).view(cls_result.size(0), 1).expand(cls_result.size(0), 1)).squeeze(1)
result_num = num_from_bins(bin, torch.argmax(cls_result), result)
gt_num = num_from_bins(bin, cls_gt, reg_gt)
return torch.abs(result_num - gt_num)
# basis = basis_from_ori(torch.rand(2))
# coeffs = torch.rand((2, 3))
# centroid = torch.rand((2, 3))
# get_corners_of_bb3d(basis, coeffs, centroid)