-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtest.py
412 lines (343 loc) · 18.8 KB
/
test.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
"""
This file belongs to the MultiBodySync code repository and is distributed for free.
Author: Jiahui Huang <[email protected]>
"""
import yaml
import tqdm
import argparse
import open3d as o3d
import torch
import torch.nn as nn
import numpy as np
from pathlib import Path
from utils.nn_util import break_leading_dim, knead_leading_dims
from torch.utils.data import DataLoader
from utils import pointnet2_util
from models.flow_net import FlowNet
from models.conf_net import ConfNet, get_network_input
from models.mot_net import MotNet
from models.full_net import compose_dense, apply_laplacian, perm_to_symm_flow, feature_propagation
from utils.sync_util import sync_perm, sync_motion_seg, motion_synchronization_spectral, fit_motion_svd_batch
from dataset import MultibodyDataset, DatasetSpec as ds
# Borrowed from PointGroup
COLOR20 = np.array(
[[230, 25, 75], [60, 180, 75], [255, 225, 25], [0, 130, 200], [245, 130, 48],
[145, 30, 180], [70, 240, 240], [240, 50, 230], [210, 245, 60], [250, 190, 190],
[0, 128, 128], [230, 190, 255], [170, 110, 40], [255, 250, 200], [128, 0, 0],
[170, 255, 195], [128, 128, 0], [255, 215, 180], [0, 0, 128], [128, 128, 128]])
def build_pointcloud(pc, cid: np.ndarray = None):
assert pc.shape[1] == 3 and len(pc.shape) == 2, f"Point cloud is of size {pc.shape} and cannot be displayed!"
point_cloud = o3d.geometry.PointCloud()
point_cloud.points = o3d.utility.Vector3dVector(pc)
if cid is not None:
assert cid.shape[0] == pc.shape[0], f"Point and color id must have same size {cid.shape[0]}, {pc.shape[0]}"
assert cid.ndim == 1, f"color id must be of size (N,) currently ndim = {cid.ndim}"
point_cloud.colors = o3d.utility.Vector3dVector(COLOR20[cid % COLOR20.shape[0]] / 255.)
return point_cloud
def binarize_motion(mat: torch.Tensor):
n_batch, _, K, N = mat.size()
while True:
mat_bin = torch.zeros_like(mat)
amax_ind = mat.argmax(dim=1, keepdim=True)
mat_bin.scatter_(dim=1, index=amax_ind, value=1.)
point_count = torch.sum(mat_bin, dim=-1)
valid_Bs = torch.all(knead_leading_dims(2, point_count) > 2, dim=-1)
if torch.all(valid_Bs):
break
mat = knead_leading_dims(2, mat)[valid_Bs]
mat = mat.view(n_batch, -1, K, N)
return mat_bin
def remove_motion_outliers(motion_absolute, xyz):
"""
:param motion_absolute: (B, [s, K, N)
:param xyz: (B, [K, N)
"""
NB_CNT = 10
NB_THRES = 4
from sklearn.neighbors import NearestNeighbors
import scipy.stats
motion_absolute, xyz = motion_absolute[0], xyz[0]
new_motion = torch.zeros_like(motion_absolute)
segm = torch.argmax(motion_absolute, dim=0) # (K, N)
n_view = xyz.size(0)
n_point = xyz.size(1)
for view_i in range(n_view):
xyz_i = xyz[view_i].cpu().numpy()
segm_i = segm[view_i].cpu().numpy()
nbrs = NearestNeighbors(n_neighbors=NB_CNT).fit(xyz_i)
nb_inds = nbrs.kneighbors(xyz_i, return_distance=False) # (N, 10)
nb_segm = segm_i[nb_inds] # (N, 10)
gcount = np.sum(nb_segm == segm_i[:, np.newaxis], axis=-1) # (N, )
gsegm = scipy.stats.mode(nb_segm, axis=1).mode[:, 0] # (N, )
segm_i[gcount < NB_THRES] = gsegm[gcount < NB_THRES]
new_motion[segm_i, view_i, np.arange(n_point)] = 1
return new_motion
def perform_icp(xyz, v_base: int, R_init, t_init, segm):
xyz, R_init, t_init, segm = xyz[0], R_init[0], t_init[0], segm[0]
segm = segm.bool()
n_view = xyz.size(0)
n_point = xyz.size(1)
n_body = R_init.size(0)
all_Rs = []
all_ts = []
for body_i in range(n_body):
for view_i in range(n_view):
R_vi = R_init[body_i, view_i].cpu().numpy()
t_vi = t_init[body_i, view_i].cpu().numpy()
T_vi = np.identity(4)
T_vi[:3, :3] = R_vi
T_vi[:3, 3] = t_vi
xyz_v = build_pointcloud(xyz[v_base, segm[body_i, v_base]].cpu().numpy())
xyz_i = build_pointcloud(xyz[view_i, segm[body_i, view_i]].cpu().numpy())
new_T_vi = o3d.pipelines.registration.registration_icp(
xyz_i, xyz_v, 0.25, T_vi,
o3d.pipelines.registration.TransformationEstimationPointToPoint(),
o3d.pipelines.registration.ICPConvergenceCriteria(max_iteration=1000))
new_T_vi = new_T_vi.transformation
all_Rs.append(new_T_vi[:3, :3])
all_ts.append(new_T_vi[:3, 3])
all_Rs = torch.from_numpy(np.stack(all_Rs, axis=0)).float().cuda().reshape(1, n_body, n_view, 3, 3)
all_ts = torch.from_numpy(np.stack(all_ts, axis=0)).float().cuda().reshape(1, n_body, n_view, 3)
return all_Rs, all_ts
def check_connectivity(mat: torch.Tensor):
mat = mat + torch.eye(mat.size(1), device=mat.device, dtype=mat.dtype).unsqueeze(0)
now_mat = mat
for vi in range(1, mat.size(1)):
now_mat = torch.bmm(now_mat, mat.transpose(-1, -2))
return torch.all(torch.all(now_mat > 0, dim=-1), dim=-1)
def edge_prune(w: torch.Tensor, th_list: list):
for weight_th in th_list:
weight_mask = (w > weight_th).float()
if torch.all(check_connectivity(weight_mask)):
w = w * weight_mask # (B, K, K)
break
return w
class TestTimeFullNet(nn.Module):
def __init__(self, args):
super().__init__()
self.args = args
self.n_iter = 4
self.rigid_n_iter = 1
self.t = 0.01
self.nsample_motion = 256
self.mcut_thres = args.alpha
assert self.rigid_n_iter <= self.n_iter
self.flow_net = FlowNet()
self.conf_net = ConfNet()
self.mot_net = MotNet()
def symm_flow_to_perm(self, pc1: torch.Tensor, flow: torch.Tensor, pc2: torch.Tensor, weight: torch.Tensor):
n_point = pc1.size(1)
dist12 = -torch.cdist(pc1 + flow[:, 0], pc2)
dist21 = -torch.cdist(pc1, pc2 + flow[:, 1])
weight_mat = torch.stack([weight[:, 0].unsqueeze(-1).repeat(1, 1, n_point),
weight[:, 1].unsqueeze(-2).repeat(1, n_point, 1)])
weight_mat /= torch.sum(weight_mat, dim=0, keepdim=True)
dist = torch.nn.functional.softmax((dist12 * weight_mat[0] + dist21 * weight_mat[1]) / self.t,
-1)
return dist
def forward(self, xyz):
"""
:param xyz: (1, K, N, 3)
"""
n_batch, n_view, n_point, _ = xyz.size()
assert n_batch == 1, "Test time algorithm only supports batch size 1"
# Sub-Sample for motion synchronization.
xyz_gathered = xyz.reshape(n_batch * n_view, n_point, 3).contiguous()
sub_inds = pointnet2_util.furthest_point_sample(xyz_gathered, self.nsample_motion).long()
xyz_down = pointnet2_util.gather_nd(xyz_gathered, sub_inds)
sub_inds = break_leading_dim([n_batch, n_view], sub_inds)
flow_init = None
xyz_transformed = xyz
motion_absolute = None
for iter_i in range(self.n_iter):
# Infer Flow & Pair-wise flow weight
perm_dict = {}
weight_bin_dict = {}
for view_i in range(n_view):
for view_j in range(view_i + 1, n_view):
pc_i = xyz_transformed[:, view_i]
pc_j = xyz_transformed[:, view_j]
# PD-Flow
flow_ij, _, _, _, _ = self.flow_net.forward(pc_i, pc_j, pc_i, pc_j)
flow_ji, _, _, _, _ = self.flow_net.forward(pc_j, pc_i, pc_j, pc_i)
flow_ij = flow_ij[0].transpose(-1, -2)
flow_ji = flow_ji[0].transpose(-1, -2)
if flow_init is not None:
flow_init_i_jjt = feature_propagation(pc_i + flow_ij, pc_j, flow_init[:, view_j], False)
flow_ij = flow_ij + flow_init[:, view_i] - flow_init_i_jjt
flow_init_j_iit = feature_propagation(pc_j + flow_ji, pc_i, flow_init[:, view_i], False)
flow_ji = flow_ji + flow_init[:, view_j] - flow_init_j_iit
flow_ij = torch.stack([flow_ij, flow_ji], dim=1)
_, weight_ij = self.conf_net(get_network_input(xyz[:, view_i], xyz[:, view_j], flow_ij[:, 0]))
_, weight_ji = self.conf_net(get_network_input(xyz[:, view_j], xyz[:, view_i], flow_ij[:, 1]))
weight_ij = torch.stack([weight_ij, weight_ji], dim=1)
weight_ij.sigmoid_()
weight_bin_score = torch.sum(weight_ij > 0.5, dim=-1).float() / n_point
weight_bin_score = torch.mean(weight_bin_score, dim=-1)
weight_bin_score = weight_bin_score.reshape(n_batch, 1, 1).float()
perm_ij = self.symm_flow_to_perm(xyz[:, view_i], flow_ij, xyz[:, view_j], weight_ij)
perm_dict[(view_i, view_j)] = perm_ij
weight_bin_dict[(view_i, view_j)] = weight_bin_score
# Synchronize permutation and get refined flow
perm_dense = compose_dense(perm_dict, n_view, torch.eye(n_point).cuda().unsqueeze(0).repeat(n_batch, 1, 1))
weight_dense = compose_dense(weight_bin_dict, n_view, torch.zeros_like(weight_bin_dict[(0, 1)]))
weight_dense = weight_dense + weight_dense.transpose(-1, -2)
# Determine the weight based on connectivity
weight_dense = edge_prune(weight_dense, [0.75, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.0])
perm_dense = apply_laplacian(perm_dense, n_view, weight_dense)
perm_dense = sync_perm(perm_dense, n_point, 1.0e-4)
# Infer Segmentation.
flow_dict = {}
motion_dict = {}
sync_weight_dict = {}
motion_canonical_scale = None
for view_i in range(n_view):
for view_j in range(view_i + 1, n_view):
perm_ij = perm_dense[:, view_i * n_point: (view_i + 1) * n_point,
view_j * n_point: (view_j + 1) * n_point] # (B, N, N)
flow_ij = perm_to_symm_flow(xyz[:, view_i], xyz[:, view_j],
perm_ij / (self.t ** 2)) # (B, 2, N, 3)
# Cache flow:
flow_dict[(view_i, view_j)] = flow_ij
_, motion_ij, _, _ = self.mot_net(xyz[:, [view_i, view_j]], flow_ij, sub_inds[:, [view_i, view_j]])
motion_ij = motion_ij.sigmoid()
motion_ij_values = motion_ij.view(motion_ij.size(0), -1)
motion_ij_scale = torch.mean(motion_ij_values, dim=-1, keepdim=True).unsqueeze(-1)
if view_i == 0 and view_j == 1:
motion_canonical_scale = motion_ij_scale
else:
motion_ij = motion_ij / motion_ij_scale * motion_canonical_scale
motion_ij = torch.clamp(motion_ij, 0.0, 1.0)
motion_dict[(view_i, view_j)] = motion_ij
# Re-evaluate the flow weights, to be used in transformation estimation.
_, weight_ij = self.conf_net(get_network_input(xyz[:, view_i], xyz[:, view_j], flow_ij[:, 0]))
_, weight_ji = self.conf_net(get_network_input(xyz[:, view_j], xyz[:, view_i], flow_ij[:, 1]))
weight_ij = torch.stack([weight_ij, weight_ji], dim=1)
sync_weight_dict[(view_i, view_j)] = weight_ij.sigmoid()
# Pairwise motion synchronization
motion_dense = compose_dense(motion_dict, n_view, torch.zeros_like(motion_dict[(0, 1)]))
motion_absolute = sync_motion_seg(motion_dense, t=0.0, cut_thres=self.mcut_thres)
sync_s = motion_absolute.size(-1)
motion_absolute /= motion_absolute.sum(-1, keepdim=True)
motion_absolute = feature_propagation(
xyz_gathered, xyz_down,
motion_absolute.reshape(n_batch * n_view, self.nsample_motion, sync_s).transpose(-1, -2)).reshape(
n_batch, n_view, sync_s, -1).permute(0, 2, 1, 3)
# Binarize segmentation at test time.
motion_absolute = binarize_motion(motion_absolute)
sync_s = motion_absolute.size(1)
motion_absolute = remove_motion_outliers(motion_absolute, xyz)
tmp_s = 1 if iter_i < self.rigid_n_iter else sync_s
motion_absolute = motion_absolute.view(n_batch * sync_s, n_view, n_point)
R_list = []
t_list = []
w_list = []
for view_i in range(n_view):
R_sub_list = []
t_sub_list = []
w_sub_list = []
for view_j in range(n_view):
if view_i < view_j:
flow_ij = flow_dict[(view_i, view_j)][:, 0]
weight_ij = sync_weight_dict[(view_i, view_j)][:, 0]
elif view_i > view_j:
flow_ij = flow_dict[(view_j, view_i)][:, 1]
weight_ij = sync_weight_dict[(view_j, view_i)][:, 1]
else:
R_sub_list.append(torch.eye(3).cuda().unsqueeze(0).repeat(n_batch * tmp_s, 1, 1))
t_sub_list.append(torch.zeros(n_batch * tmp_s, 3).cuda())
w_sub_list.append(torch.zeros(n_batch * tmp_s, ).cuda())
continue
flow_ij = flow_ij.unsqueeze(1).repeat(1, tmp_s, 1, 1)
flow_ij = knead_leading_dims(2, flow_ij)
weight_ij = weight_ij.unsqueeze(1).repeat(1, tmp_s, 1)
weight_ij = knead_leading_dims(2, weight_ij)
xyz_i = xyz[:, view_i].unsqueeze(1).repeat(1, tmp_s, 1, 1)
xyz_i = knead_leading_dims(2, xyz_i)
if tmp_s == 1:
R_ij, t_ij = fit_motion_svd_batch(xyz_i,
xyz_i + flow_ij, weight_ij)
w_ij = torch.mean(weight_ij, -1)
else:
local_weight = motion_absolute[:, view_i] * weight_ij
R_ij, t_ij = fit_motion_svd_batch(xyz_i,
xyz_i + flow_ij,
local_weight)
t_ij.clamp_(-2.0, 2.0)
w_ij = torch.sum(local_weight, -1) / torch.sum(motion_absolute[:, view_i], -1)
R_sub_list.append(R_ij)
t_sub_list.append(t_ij)
w_sub_list.append(w_ij)
R_list.append(R_sub_list)
t_list.append(t_sub_list)
w_list.append(w_sub_list)
motion_absolute = break_leading_dim([n_batch, sync_s], motion_absolute)
trans_global_weight = torch.stack([torch.stack(sl, -1) for sl in w_list], 1)
trans_global_weight = (trans_global_weight + trans_global_weight.transpose(-1, -2)) / 2.
trans_global_weight = edge_prune(trans_global_weight, [0.8, 0.75, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.0])
# This motion sync part seems not contribute much to the accuracy at test time.
# So we disable it by default. You can always re-enable it.
# R_sync, t_sync = motion_synchronization_spectral(R_list, t_list,
# trans_global_weight,
# fallback_on_error=True)
R_sync = torch.stack(R_list[0], dim=1)
t_sync = torch.stack(t_list[0], dim=1)
R_sync = R_sync.reshape(n_batch, tmp_s, n_view, 3, 3)
t_sync = t_sync.reshape(n_batch, tmp_s, n_view, 3)
# Enforce Gauge-freedom to be aligned with the first-view. (s.t. all K == 0 will be identity)
# So the transformed point cloud do not have weird shapes, which will otherwise make flow bad.
v_base = n_view - 1
R_base = R_sync[:, :, v_base].transpose(-1, -2)
t_base = -torch.einsum('bsij,bsj->bsi', R_base, t_sync[:, :, v_base])
R_sync_cal = torch.einsum('bskij,bsjm->bskim', R_sync, R_base)
t_sync_cal = torch.einsum('bskij,bsj->bski', R_sync, t_base) + t_sync
R_sync, t_sync = R_sync_cal, t_sync_cal
R_sync = R_sync.transpose(-1, -2)
t_sync = -torch.einsum('bskij,bskj->bski', R_sync, t_sync)
# Ignore some outliers.
t_sync[t_sync > 1.0] = 0.0
if iter_i >= self.rigid_n_iter:
R_sync, t_sync = perform_icp(xyz, v_base, R_sync, t_sync, motion_absolute)
# Apply the motion to the points, so that for next iteration
# Flow will be made easier.
xyz_transformed = torch.einsum('bskij,bknj->bskni', R_sync, xyz) + \
t_sync.unsqueeze(-2)
if tmp_s == 1:
xyz_transformed = xyz_transformed.squeeze(1)
else:
xyz_transformed = torch.einsum('bskni,bskn->bkni', xyz_transformed, motion_absolute)
xyz_transformed = xyz_transformed.contiguous()
flow_init = xyz_transformed - xyz
motion_absolute = motion_absolute.permute(0, 2, 3, 1)
return motion_absolute, xyz_transformed
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('config', type=str, help='Config files')
# Read parameters
args = parser.parse_args()
with Path(args.config).open() as f:
configs = yaml.load(f, Loader=yaml.FullLoader)
for ckey, cvalue in configs.items():
args.__dict__[ckey] = cvalue
assert args.type == "full", "Your config file must be of type 'full'."
# Modify the following to test on your own data.
test_set = MultibodyDataset(args.test_base_folder, [ds.PC, ds.SEGM, ds.FULL_FLOW], 'test', None)
test_loader = DataLoader(test_set, batch_size=1, shuffle=False, num_workers=1, pin_memory=True)
weight_path = args.save_path + "/best.pth.tar"
model = TestTimeFullNet(args)
model.load_state_dict(torch.load(weight_path)['model_state'])
model.cuda().eval()
all_ious = []
with tqdm.tqdm(enumerate(test_loader, 0), total=len(test_loader), desc='test') as tbar:
for i, data in tbar:
with torch.no_grad():
inputs, segm, full_flow = data
inputs = inputs.cuda()
segm = segm.cuda()
pd_segm, _ = model(inputs)
segmented_pcds = []
for view_i in range(inputs.size(1)):
segmented_pcds.append(build_pointcloud(inputs[0, view_i].cpu().numpy(),
pd_segm.argmax(-1)[0, view_i].cpu().numpy()))
segmented_pcds[-1].translate([view_i, 0.0, 0.0])
o3d.visualization.draw_geometries(segmented_pcds)