Repository navigation
Expand file tree
/
Copy pathmetric.py
More file actions
333 lines (261 loc) · 11.1 KB
/
Copy pathmetric.py
File metadata and controls
333 lines (261 loc) · 11.1 KB
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
import numpy as np
from scipy.spatial.transform import Rotation
import math
import torch
import src.utils as utils
from src.util.lie_metric.MSEE import SE3_to_se3, cal_metric
def quat_to_rotmat_batch(q):
"""
Args:
q: [B,4]
Returns:
rotmat: [B,3,3]
"""
qw = q[:, 0]
qx = q[:, 1]
qy = q[:, 2]
qz = q[:, 3]
r00 = 1 - 2 * qy ** 2 - 2 * qz ** 2
r01 = 2 * qx * qy - 2 * qz * qw
r02 = 2 * qx * qz + 2 * qy * qw
r10 = 2 * qx * qy + 2 * qz * qw
r11 = 1 - 2 * qx ** 2 - 2 * qz ** 2
r12 = 2 * qy * qz - 2 * qx * qw
r20 = 2 * qx * qz - 2 * qy * qw
r21 = 2 * qy * qz + 2 * qx * qw
r22 = 1 - 2 * qx ** 2 - 2 * qy ** 2
rotmat = np.stack([r00, r01, r02, r10, r11, r12, r20, r21, r22], axis=-1)
return rotmat.reshape(-1, 3, 3)
def mult_extrinsic_batch(m1, m2):
"""
Args:
m1: [B,3,4]
m2: [B,3,4]
Returns:
mult: [B,3,4]
"""
B = m1.shape[0]
padding = np.array([0, 0, 0, 1]).reshape(1, 1, 4).repeat(B, axis=0)
m1 = np.concatenate([m1, padding], axis=-2) # B,4,4
m2 = np.concatenate([m2, padding], axis=-2) # B,4,4
mult = m1 @ m2
return mult[:, :3, :] # B,3,4
def inv_extrinsic(m):
B = m.shape[0]
padding = np.array([0, 0, 0, 1]).reshape(1, 1, 4).repeat(B, axis=0)
m1 = np.concatenate([m, padding], axis=-2) # B,4,4
return np.linalg.inv(m1)[:, :3, :]
def rotmat_to_euler(rotmat, out='rad'):
"""
Args:
rotmat: [B,3,3]
out: "rad" or "deg"
"""
sy = np.sqrt(rotmat[:, 0, 0] * rotmat[:, 0, 0] + rotmat[:, 1, 0] * rotmat[:, 1, 0])
singular = sy < 1e-6 # B
roll = np.empty_like(sy)
pitch = np.empty_like(sy)
yaw = np.empty_like(sy)
nonsingular = np.logical_not(singular)
roll[nonsingular] = np.arctan2(rotmat[nonsingular, 2, 1], rotmat[nonsingular, 2, 2])
pitch[nonsingular] = np.arctan2(-rotmat[nonsingular, 2, 0], sy[nonsingular])
yaw[nonsingular] = np.arctan2(rotmat[nonsingular, 1, 0], rotmat[nonsingular, 0, 0])
roll[singular] = np.arctan2(-rotmat[singular, 1, 2], rotmat[singular, 1, 1])
pitch[singular] = np.arctan2(-rotmat[singular, 2, 0], sy[singular])
yaw[singular] = 0
if out == 'rad':
return np.stack([roll, pitch, yaw], -1)
elif out == 'deg':
return np.stack([roll, pitch, yaw], -1) * 180 / np.pi
def calibration_error_batch(e1, e2):
# new setting E_{pred}^{-1} Pr^{-1}<=>E_{gt}
inv_e1 = inv_extrinsic(e1)
err_e = mult_extrinsic_batch(inv_e1, e2)
err_rotmat = err_e[:, :3, :3]
err_trans = err_e[:, :3, 3]
err_euler = rotmat_to_euler(err_rotmat, out='deg')
err_euler = np.abs(err_euler)
err_trans = np.abs(err_trans)
return err_euler[:, 0], err_euler[:, 1], err_euler[:, 2], \
err_trans[:, 0], err_trans[:, 1], err_trans[:, 2]
def getExtrinsic(out3, data_valid,out_raw=False):
pred_decalib_quat_real = out3[:, :4].cpu().detach().numpy()
pred_decalib_quat_dual = out3[:, 4:].cpu().detach().numpy().reshape(-1, 3, 1)
gt_decalib_quat_real = data_valid['decalib_real_gt'].numpy()
gt_decalib_quat_dual = data_valid['decalib_dual_gt'].numpy().reshape(-1, 3, 1)
init_extrinsic = data_valid['init_extrinsic']
pred_decalib_rot = quat_to_rotmat_batch(pred_decalib_quat_real) # [B,3,3]
# [B,3,4]
pred_decalib_extrinsic = np.concatenate([pred_decalib_rot, pred_decalib_quat_dual], axis=-1)
pred_extrinsic = mult_extrinsic_batch(pred_decalib_extrinsic, init_extrinsic) # [B,3,4]
gt_decalib_rot = quat_to_rotmat_batch(gt_decalib_quat_real)
gt_decalib_extrinsic = np.concatenate([gt_decalib_rot, gt_decalib_quat_dual], axis=-1)
gt_extrinsic = mult_extrinsic_batch(gt_decalib_extrinsic, init_extrinsic)
if out_raw:
return pred_extrinsic, gt_extrinsic,pred_decalib_extrinsic,gt_decalib_extrinsic
else:
return pred_extrinsic, gt_extrinsic
def cal_rete_once(out3, data_valid):
with torch.no_grad():
pred_decalib_quat_real = out3[:, :4].cpu().detach().numpy()
pred_decalib_quat_dual = out3[:, 4:].cpu().detach().numpy().reshape(-1, 3, 1)
gt_decalib_quat_real = data_valid['decalib_real_gt'].numpy()
gt_decalib_quat_dual = data_valid['decalib_dual_gt'].numpy().reshape(-1, 3, 1)
pred_decalib_rot = quat_to_rotmat_batch(pred_decalib_quat_real) # [B,3,3]
# [B,3,4]
pred_raw = np.concatenate([pred_decalib_rot, pred_decalib_quat_dual], axis=-1)
gt_decalib_rot = quat_to_rotmat_batch(gt_decalib_quat_real)
gt_raw = np.concatenate([gt_decalib_rot, gt_decalib_quat_dual], axis=-1)
P_diff = mult_extrinsic_batch(inv_extrinsic(pred_raw), gt_raw)
t_diff = np.linalg.norm(P_diff[:, :3, 3], 2, -1)
r_diff = P_diff[:, :3, :3]
R_diff = Rotation.from_matrix(r_diff)
angles_diff = np.sum(np.abs(R_diff.as_euler('xzy', degrees=True)), -1)
return angles_diff.mean(),t_diff.mean()
def getExtrinsicRt(R, t, data_valid):
gt_decalib_quat_real = data_valid['decalib_real_gt'].numpy()
gt_decalib_quat_dual = data_valid['decalib_dual_gt'].numpy().reshape(-1, 3, 1)
init_extrinsic = data_valid['init_extrinsic']
# [B,3,4]
pred_decalib_extrinsic = np.concatenate([R.detach().cpu().numpy(),
t.unsqueeze(-1).detach().cpu().numpy()], axis=-1)
pred_extrinsic = mult_extrinsic_batch(pred_decalib_extrinsic, init_extrinsic) # [B,3,4]
gt_decalib_rot = quat_to_rotmat_batch(gt_decalib_quat_real)
gt_decalib_extrinsic = np.concatenate([gt_decalib_rot, gt_decalib_quat_dual], axis=-1)
gt_extrinsic = mult_extrinsic_batch(gt_decalib_extrinsic, init_extrinsic)
return pred_extrinsic, gt_extrinsic
def eval_acc(pm, intrinsic, img_size, decalib_q, decalib_t, sigmoid=False):
if len(pm) == 2:
l_gt = utils.get_projection_gt(pm[1], intrinsic, img_size, decalib_q, decalib_t)
else:
l_gt = pm[2].argmax(-1).long()
pred = pm[0].argmax(-1) if not sigmoid else torch.ge(pm[0].squeeze(-1), 0.5).long()
N = pred.shape[1]
err = torch.sum(torch.abs(pred - l_gt), -1)
acc = (1 - err.float() / N).cpu().numpy()
return acc
def eval_msee(out3, gt_se3):
# TODO: MRR
# pred [B,3,4] gt_se3 [B,6]
pred_decalib_quat_real = out3[:, :4].cpu().detach().numpy()
pred_decalib_quat_dual = out3[:, 4:].cpu().detach().numpy().reshape(-1, 3, 1)
pred_decalib_rot = quat_to_rotmat_batch(pred_decalib_quat_real) # [B,3,3]
# [B,3,4]
pred = np.concatenate([pred_decalib_rot, pred_decalib_quat_dual], axis=-1)
pred_se3 = np.stack([SE3_to_se3(np.concatenate([pred[i], np.array([[0, 0, 0, 1]])], axis=0))
for i in range(pred.shape[0])], axis=0) # B,6
msee = cal_metric(pred_se3, gt_se3)
return msee
def eval_mrr(msee, gt_se3):
se3_noise = cal_metric(np.zeros_like(np.array(gt_se3)), gt_se3)
rr = 1 - msee / se3_noise
return rr
class RteRreEval(object):
def __init__(self, threshold=False, rre_th=10., rte_th=5.):
self.t_diff = []
self.r_diff = []
self.t_diff_all = []
self.r_diff_all = []
self.threshold = threshold
self.rre_th = rre_th
self.rte_th = rte_th
self.acc_count = 0
self.all_count = 0
def reset(self):
self.t_diff.clear()
self.r_diff.clear()
self.acc_count = 0
self.all_count = 0
def get_recall(self):
return self.acc_count / self.all_count
def addBatch(self, pred_extrinsic, gt_extrinsic):
"""
Args:
pred_extrinsic: [B,3,4]
gt_extrinsic: [B,3,4]
Returns:
"""
# change inv(pred)@gt to pred@inv(gt) for gt:(Pc) init: (Pr@Pc)=>pred@Pr
# P_diff = mult_extrinsic_batch(pred_extrinsic, inv_extrinsic(gt_extrinsic))
P_diff = mult_extrinsic_batch(inv_extrinsic(pred_extrinsic), gt_extrinsic)
t_diff = np.linalg.norm(P_diff[:, :3, 3], 2, -1)
r_diff = P_diff[:, :3, :3]
R_diff = Rotation.from_matrix(r_diff)
angles_diff = np.sum(np.abs(R_diff.as_euler('xzy', degrees=True)), -1)
self.all_count += len(angles_diff)
if self.threshold:
mask = np.logical_and(t_diff < self.rte_th, angles_diff < self.rre_th)
masked_t_diff = t_diff[mask]
masked_angles_diff = angles_diff[mask]
self.acc_count += len(masked_t_diff)
self.t_diff.extend(list(masked_t_diff))
self.r_diff.extend(list(masked_angles_diff))
else:
self.acc_count += len(angles_diff)
self.t_diff.extend(list(t_diff))
self.r_diff.extend(list(angles_diff))
self.t_diff_all.extend(list(t_diff))
self.r_diff_all.extend(list(angles_diff))
return list(angles_diff), list(t_diff)
def evalSeq(self):
t_diff = np.array(self.t_diff)
r_diff = np.array(self.r_diff)
rte_sigma = math.sqrt(np.var(t_diff))
rre_sigma = math.sqrt(np.var(r_diff))
rte_mean = t_diff.mean()
rre_mean = r_diff.mean()
return rte_mean, rte_sigma, rre_mean, rre_sigma
def save_metric(self, path):
np.savez(path, RRE=np.array(self.r_diff_all),
RTE=np.array(self.t_diff_all))
def quatmultiply(q, r, device='cpu'):
"""
Batch quaternion multiplication
Args:
q (torch.Tensor/np.ndarray): shape=[Nx4]
r (torch.Tensor/np.ndarray): shape=[Nx4]
device (str): 'cuda' or 'cpu'
Returns:
torch.Tensor: shape=[Nx4]
"""
if isinstance(q, torch.Tensor):
t = torch.zeros(q.shape[0], 4, device=device)
elif isinstance(q, np.ndarray):
#print(q.shape[0])
t = np.zeros((q.shape[0], 4))
else:
raise TypeError("Type not supported")
t[:, 0] = r[:, 0] * q[:, 0] - r[:, 1] * q[:, 1] - r[:, 2] * q[:, 2] - r[:, 3] * q[:, 3]
t[:, 1] = r[:, 0] * q[:, 1] + r[:, 1] * q[:, 0] - r[:, 2] * q[:, 3] + r[:, 3] * q[:, 2]
t[:, 2] = r[:, 0] * q[:, 2] + r[:, 1] * q[:, 3] + r[:, 2] * q[:, 0] - r[:, 3] * q[:, 1]
t[:, 3] = r[:, 0] * q[:, 3] - r[:, 1] * q[:, 2] + r[:, 2] * q[:, 1] + r[:, 3] * q[:, 0]
return t
def quatinv(q):
"""
Batch quaternion inversion
Args:
q (torch.Tensor/np.ndarray): shape=[Nx4]
Returns:
torch.Tensor/np.ndarray: shape=[Nx4]
"""
if isinstance(q, torch.Tensor):
t = q.clone()
elif isinstance(q, np.ndarray):
t = q.copy()
else:
raise TypeError("Type not supported")
t *= -1
t[:, 0] *= -1
return t
def quaternion_distance(q, r):
"""
Batch quaternion distances, used as loss
Args:
q (torch.Tensor): shape=[Nx4]
r (torch.Tensor): shape=[Nx4]
device (str): 'cuda' or 'cpu'
Returns:
torch.Tensor: shape=[N]
"""
t = quatmultiply(q, quatinv(r))
return 2 * np.arctan2(np.linalg.norm(t[:, 1:], axis=1), np.abs(t[:, 0]))