/
tester.py
297 lines (212 loc) · 10.3 KB
/
tester.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
from lib.trainer import Trainer
import torch
from tqdm import tqdm
from models.loss import MatchMotionLoss as MML
import numpy as np
from models.matching import Matching as CM
import math
class _3DMatchTester(Trainer):
"""
3DMatch tester
"""
def __init__(self,args):
Trainer.__init__(self, args)
def test(self):
n = 3
afmr = 0.
arr = 0
air = 0
for i in range(n): # combat ransac nondeterministic
thr =0.05
rr, ir, fmr = self.test_thr(thr)
afmr+=fmr
arr+=rr
air+=ir
print( "conf_threshold", thr, "registration recall:", rr, " Inlier rate:", ir, "FMR:", fmr)
print("average registration recall:", arr / n, afmr/n, air/n)
# print ("registration recall:", self.test_thr())
def test_thr(self, conf_threshold=None):
# print('Start to evaluate on test datasets...')
# os.makedirs(f'{self.snapshot_dir}/{self.config.dataset}',exist_ok=True)
num_iter = math.ceil(len(self.loader['test'].dataset) // self.loader['test'].batch_size)
c_loader_iter = self.loader['test'].__iter__()
self.model.eval()
success1 = 0.
IR=0.
FMR=0.
with torch.no_grad():
for idx in tqdm(range(num_iter)): # loop through this epoch
##################################
if self.timers: self.timers.tic('load batch')
inputs = c_loader_iter.next()
for k, v in inputs.items():
if type(v) == list:
inputs[k] = [item.to(self.device) for item in v]
elif type(v) in [dict, float, type(None), np.ndarray]:
pass
else:
inputs[k] = v.to(self.device)
if self.timers: self.timers.toc('load batch')
##################################
if self.timers: self.timers.tic('forward pass')
data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2]
if self.timers: self.timers.toc('forward pass')
match_pred, _, _ = CM.get_match(data['conf_matrix_pred'], thr=conf_threshold, mutual=False)
rot, trn = MML.ransac_regist_coarse(data['s_pcd'], data['t_pcd'], data['src_mask'], data['tgt_mask'], match_pred)
ir = MML.compute_inlier_ratio(match_pred, data, inlier_thr=0.1).mean()
rr1 = MML.compute_registration_recall(rot, trn, data, thr=0.2) # 0.2m
vis = False
if vis:
pcd = data['points'][0].cpu().numpy()
lenth = data['stack_lengths'][0][0]
spcd, tpcd = pcd[:lenth] , pcd[lenth:]
import mayavi.mlab as mlab
c_red = (224. / 255., 0 / 255., 125 / 255.)
c_pink = (224. / 255., 75. / 255., 232. / 255.)
c_blue = (0. / 255., 0. / 255., 255. / 255.)
scale_factor = 0.02
# mlab.points3d(s_pc[ :, 0] , s_pc[ :, 1], s_pc[:, 2], scale_factor=scale_factor , color=c_blue)
mlab.points3d(spcd[:, 0], spcd[:, 1], spcd[:, 2], scale_factor=scale_factor,
color=c_red)
mlab.points3d(tpcd[:, 0], tpcd[:, 1], tpcd[:, 2], scale_factor=scale_factor,
color=c_blue)
mlab.show()
spcd = ( np.matmul(rot, spcd.T) + trn ).T
mlab.points3d(spcd[:, 0], spcd[:, 1], spcd[:, 2], scale_factor=scale_factor,
color=c_red)
mlab.points3d(tpcd[:, 0], tpcd[:, 1], tpcd[:, 2], scale_factor=scale_factor,
color=c_blue)
mlab.show()
bs = len(rot)
assert bs==1
success1 += bs * rr1
IR += bs*ir
FMR += (ir>0.05).float()
recall1 = success1/len(self.loader['test'].dataset)
IRate = IR/len(self.loader['test'].dataset)
FMR = FMR/len(self.loader['test'].dataset)
return recall1, IRate, FMR
def blend_anchor_motion (query_loc, reference_loc, reference_flow , knn=3, search_radius=0.1) :
'''approximate flow on query points
this function assume query points are sub- or un-sampled from reference locations
@param query_loc:[m,3]
@param reference_loc:[n,3]
@param reference_flow:[n,3]
@param knn:
@return:
blended_flow:[m,3]
'''
from datasets.utils import knn_point_np
dists, idx = knn_point_np (knn, reference_loc, query_loc)
dists[dists < 1e-10] = 1e-10
mask = dists>search_radius
dists[mask] = 1e+10
weight = 1.0 / dists
weight = weight / np.sum(weight, -1, keepdims=True) # [B,N,3]
blended_flow = np.sum (reference_flow [idx] * weight.reshape ([-1, knn, 1]), axis=1, keepdims=False)
mask = mask.sum(axis=1)<3
return blended_flow, mask
def compute_nrfmr( match_pred, data, recall_thr=0.04):
s_pcd, t_pcd = data['s_pcd'], data['t_pcd']
s_pcd_raw = data ['src_pcd_list']
sflow_list = data['sflow_list']
metric_index_list = data['metric_index_list']
batched_rot = data['batched_rot'] # B,3,3
batched_trn = data['batched_trn']
nrfmr = 0.
for i in range ( len(s_pcd_raw)):
# get the metric points' transformed position
metric_index = metric_index_list[i]
sflow = sflow_list[i]
s_pcd_raw_i = s_pcd_raw[i]
metric_pcd = s_pcd_raw_i [ metric_index ]
metric_sflow = sflow [ metric_index ]
metric_pcd_deformed = metric_pcd + metric_sflow
metric_pcd_wrapped_gt = ( torch.matmul( batched_rot[i], metric_pcd_deformed.T) + batched_trn[i] ).T
# use the match prediction as the motion anchor
match_pred_i = match_pred[ match_pred[:, 0] == i ]
s_id , t_id = match_pred_i[:,1], match_pred_i[:,2]
s_pcd_matched= s_pcd[i][s_id]
t_pcd_matched= t_pcd[i][t_id]
motion_pred = t_pcd_matched - s_pcd_matched
metric_motion_pred, valid_mask = blend_anchor_motion(
metric_pcd.cpu().numpy(), s_pcd_matched.cpu().numpy(), motion_pred.cpu().numpy(), knn=3, search_radius=0.1)
metric_pcd_wrapped_pred = metric_pcd + torch.from_numpy(metric_motion_pred).to(metric_pcd)
debug = False
if debug:
import mayavi.mlab as mlab
c_red = (224. / 255., 0 / 255., 125 / 255.)
c_pink = (224. / 255., 75. / 255., 232. / 255.)
c_blue = (0. / 255., 0. / 255., 255. / 255.)
scale_factor = 0.013
metric_pcd_wrapped_gt = metric_pcd_wrapped_gt.cpu()
metric_pcd_wrapped_pred = metric_pcd_wrapped_pred.cpu()
err = metric_pcd_wrapped_pred - metric_pcd_wrapped_gt
mlab.points3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], scale_factor=scale_factor, color=c_pink)
mlab.points3d(metric_pcd_wrapped_pred[ :, 0] , metric_pcd_wrapped_pred[ :, 1], metric_pcd_wrapped_pred[:, 2], scale_factor=scale_factor , color=c_blue)
mlab.quiver3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], err[:, 0], err[:, 1], err[:, 2],
scale_factor=1, mode='2ddash', line_width=1.)
mlab.show()
dist = torch.sqrt( torch.sum( (metric_pcd_wrapped_pred - metric_pcd_wrapped_gt)**2, dim=1 ) )
r = (dist < recall_thr).float().sum() / len(dist)
nrfmr = nrfmr + r
nrfmr = nrfmr /len(s_pcd_raw)
return nrfmr
class _4DMatchTester(Trainer):
"""
3DMatch tester
"""
def __init__(self,args):
Trainer.__init__(self, args)
def test(self):
for thr in [ 0.05, 0.1, 0.2]:
# for thr in [ 0.1 ]:
import time
start = time.time()
ir, fmr, nspl = self.test_thr(thr)
print( "conf_threshold", thr, "NFMR:", fmr, " Inlier rate:", ir, "Number sample:", nspl)
print( "time costs:", time.time() - start)
def test_thr(self, conf_threshold=None):
num_iter = math.ceil(len(self.loader['test'].dataset) // self.loader['test'].batch_size)
c_loader_iter = self.loader['test'].__iter__()
self.model.eval()
assert self.loader['test'].batch_size == 1
IR=0.
NR_FMR=0.
inlier_thr = recall_thr = 0.04
n_sample = 0.
with torch.no_grad():
for idx in tqdm(range(num_iter)): # loop through this epoch
##################################
if self.timers: self.timers.tic('load batch')
inputs = c_loader_iter.next()
for k, v in inputs.items():
if type(v) == list:
inputs[k] = [item.to(self.device) for item in v]
elif type(v) in [ dict, float, type(None), np.ndarray]:
pass
else:
inputs[k] = v.to(self.device)
if self.timers: self.timers.toc('load batch')
##################################
if self.timers: self.timers.tic('forward pass')
data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2]
if self.timers: self.timers.toc('forward pass')
match_pred, _, _ = CM.get_match(data['conf_matrix_pred'], thr=conf_threshold, mutual=True)
ir = MML.compute_inlier_ratio(match_pred, data, inlier_thr=inlier_thr, s2t_flow=data['coarse_flow'][0][None] )[0]
nrfmr = compute_nrfmr(match_pred, data, recall_thr=recall_thr)
IR += ir
NR_FMR += nrfmr
n_sample += match_pred.shape[0]
IRate = IR/len(self.loader['test'].dataset)
NR_FMR = NR_FMR/len(self.loader['test'].dataset)
n_sample = n_sample/len(self.loader['test'].dataset)
if self.timers: self.timers.print()
return IRate, NR_FMR, n_sample
def get_trainer(config):
if config.dataset == '3dmatch':
return _3DMatchTester(config)
elif config.dataset == '4dmatch':
return _4DMatchTester(config)
else:
raise NotImplementedError