-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathseg_util.py
254 lines (243 loc) · 10.7 KB
/
seg_util.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
import os
import sys
import numpy as np
from sklearn.neighbors import NearestNeighbors
import scipy.io as sio
import time
import copy
import itertools
import scipy.sparse as sparse
import scipy.sparse.csgraph as csgraph
def fit_motion(pc1, pc2):
pc1_centered = pc1-np.mean(pc1, 0, keepdims=True)
pc2_centered = pc2-np.mean(pc2, 0, keepdims=True)
R = np.matmul(np.transpose(pc1_centered), pc2_centered)
u,s,v = np.linalg.svd(R, full_matrices=True)
R = np.matmul(u,v)
if np.sum(np.abs(np.real(np.linalg.eig(R)[0])-1)<1e-5)==0:
u[:,2] = -u[:,2]
R = np.matmul(u,v)
t = np.mean(pc2-np.matmul(pc1, R), 0, keepdims=True)
return R, t
def gen_attention_mask(pc1, pcPred, pc2, segidx, segidx2, Rmodes, tmodes, nmodes, npoint, distmatrix=None, distmatrix2=None, expand_eps=0.15):
# segidx - each column corresponds to one segment
if distmatrix is None:
distmatrix = np.sqrt(np.sum((np.expand_dims(pc1, 1)-np.expand_dims(pc1,0))**2,2))
distmatrix2 = np.sqrt(np.sum((np.expand_dims(pc2, 1)-np.expand_dims(pc2,0))**2,2))
nnbr = 5
nbrs = NearestNeighbors(n_neighbors=nnbr+1, algorithm='ball_tree').fit(pc1)
_, idx = nbrs.kneighbors(pc1)
distmask = np.zeros(npoint*npoint)
for i in range(nnbr):
distmask[np.ravel_multi_index((np.arange(npoint),idx[:,i+1]),(npoint, npoint))] = 1.0
distmask = np.reshape(distmask,(npoint,npoint))
distmatrix = distmatrix*distmask
nbrs2 = NearestNeighbors(n_neighbors=nnbr+1, algorithm='ball_tree').fit(pc2)
_, idx2 = nbrs2.kneighbors(pc2)
distmask2 = np.zeros(npoint*npoint)
for i in range(nnbr):
distmask2[np.ravel_multi_index((np.arange(npoint),idx2[:,i+1]),(npoint, npoint))] = 1.0
distmask2 = np.reshape(distmask2,(npoint,npoint))
distmatrix2 = distmatrix2*distmask2
distmatrix = sparse.csr_matrix(distmatrix)
distmatrix = csgraph.floyd_warshall(distmatrix, directed=False)
distmatrix2 = sparse.csr_matrix(distmatrix2)
distmatrix2 = csgraph.floyd_warshall(distmatrix2, directed=False)
batch_pc1 = np.zeros((nmodes, npoint, 3))
for i in range(nmodes):
batch_pc1[i,:,:] = np.matmul(pc1, Rmodes[i,:,:])+tmodes[[i],:]
spatial_dist_weights = np.ones((nmodes, npoint))
for j in range(nmodes):
if np.sum(segidx[:,j])>0:
spatial_dist_weights[j,:] = np.min(distmatrix[:,segidx[:,j]>0],1)
segidx_aug = np.argmin(np.sum((batch_pc1-np.expand_dims(pcPred,0))**2,2)*spatial_dist_weights,0)+1
segidx2_aug = segidx_aug[np.argmin(np.sum((np.expand_dims(pc2,1)-np.expand_dims(pcPred,0))**2,2),1)]
pcout1 = np.zeros((nmodes, npoint, 3))
pcout2 = np.zeros((nmodes, npoint, 3))
attmask1 = np.zeros((nmodes, npoint), dtype='bool')
attmask2 = np.zeros((nmodes, npoint), dtype='bool')
for i in range(nmodes):
if np.sum(segidx[:,i])>0 or np.sum(segidx_aug==i+1)>0:
subidx1 = np.min(distmatrix[np.logical_or(segidx[:,i]>0,segidx_aug==i+1),:],0)<expand_eps
subpc1 = pc1[subidx1,:]
smppc = pc1[np.logical_or(segidx[:,i]>0,segidx_aug==i+1),:]
pcout1[i,:,:] = np.concatenate((subpc1, smppc[np.random.choice(np.arange(smppc.shape[0]),npoint-subpc1.shape[0]),:]),0)
attmask1[i,:] = subidx1
if np.sum(segidx2[:,i])>0 or np.sum(segidx2_aug==i+1)>0:
subidx2 = np.min(distmatrix2[np.logical_or(segidx2[:,i]>0,segidx2_aug==i+1),:],0)<expand_eps
subpc2 = pc2[subidx2,:]
smppc2 = pc2[np.logical_or(segidx2[:,i]>0,segidx2_aug==i+1),:]
pcout2[i,:,:] = np.concatenate((subpc2, smppc2[np.random.choice(np.arange(smppc2.shape[0]),npoint-subpc2.shape[0]),:]),0)
attmask2[i,:] = subidx2
return pcout1, pcout2, attmask1, attmask2, segidx_aug, segidx2_aug, distmatrix, distmatrix2
def motion_modes_aggreg_watt(pc1, pcPred, attmask):
# pc1, pcPred: Nmodes x Npt x 3
# attmask: Nmodes x Npt
nmode = pc1.shape[0]
npt = pc1.shape[1]
curpred_aggreg = np.zeros((npt,3,nmode))
curflownorm_aggreg = np.inf*np.ones((npt,nmode))
for i in range(nmode):
nsubpt = np.sum(attmask[i,:])
curpred_aggreg[attmask[i,:],:,i] = pcPred[i,:nsubpt,:]
curflowpred = pcPred[i,:nsubpt,:]-pc1[i,:nsubpt,:]
curflownorm_aggreg[attmask[i,:],i] = np.sqrt(np.sum(curflowpred**2,1))
modeidx = np.arange(nmode)
imx = np.argmin(curflownorm_aggreg[:,modeidx],1)
distmask = np.zeros(npt*nmode)
distmask[np.ravel_multi_index((np.arange(npt),modeidx[imx]),(npt, nmode))] = 1.0
distmask = np.reshape(distmask,(npt,nmode))
mask = np.reshape(distmask,(npt,1,nmode))
curpred_aggreg = curpred_aggreg*mask
curpred_aggreg = np.divide(np.sum(curpred_aggreg,2),np.sum(mask,2)+1e-8)
segidx = np.argmax(distmask[:,modeidx],1)
return curpred_aggreg, segidx
def decode_motion_modes(xyz, flow_pred, xyz2, trans_pred, grouping_pred, seg_pred, conf_pred, eps):
# flow_pred: npoint x 3
# trans_pred: nsmp x 18
# grouping_pred: npoint x npoint
# seg_pred: nmask x npoint
# conf_pred: nmask
npoint = seg_pred.shape[1]
nmodes = np.maximum(np.sum(conf_pred>0.5),1)
Rmodes = np.tile(np.expand_dims(np.eye(3),0),(nmodes,1,1))
tmodes = np.zeros((nmodes,3))
vidxmodes = np.ones(nmodes)
segidx = np.zeros((nmodes, npoint))
segidx2 = np.zeros(segidx.shape)
for i in range(nmodes):
if np.sum(seg_pred[i,:]>0.5)<3:
vidxmodes[i] = 0
else:
sidx = seg_pred[i,:]
R, t, inlier_idx1, inlier_idx2 = fit_motion_per_seg(xyz, xyz2, flow_pred, sidx, eps)
segidx[i,:] = inlier_idx1
segidx2[i,:] = inlier_idx2
Rmodes[i,:,:] = copy.deepcopy(R)
tmodes[i,:] = copy.deepcopy(t)
if np.sum(segidx)<5:
vidxmodes[i] = 0
nmodes = np.sum(vidxmodes).astype('int32')
Rmodes = Rmodes[vidxmodes==1,:,:]
tmodes = tmodes[vidxmodes==1,:]
segidx = segidx[vidxmodes==1,:]
segidx2 = segidx2[vidxmodes==1,:]
segidx = np.transpose(segidx,(1,0))
segidx2 = np.transpose(segidx2,(1,0))
return Rmodes, tmodes, nmodes, segidx, segidx2
def fit_motion_per_seg(pc1, pc2, flow12, segidx, eps):
# segidx: (npoint), segmentation index on pc1
# pc1: (npoint, 3)
# pc2: (npoint, 3), a partial point cloud with potential padding to form npoint
# flow12: (npoint, 3) a flow field from pc1 to pc2
th = np.maximum(np.sort(segidx)[20], 0.5)
segidx = (segidx>th).astype('float32')
maskk = 8
npt1 = pc1.shape[0]
npt2 = pc2.shape[0]
pcpred = pc1+flow12
nbrs = NearestNeighbors(n_neighbors=maskk, algorithm='ball_tree').fit(pcpred)
maskdist, maskidx = nbrs.kneighbors(pc2) # npt2 x maskk
tmpmaskidx = np.reshape(np.transpose(maskidx),maskk*npt2)
vmask = np.reshape(np.cumsum(np.reshape(segidx[tmpmaskidx],(maskk, npt2)),0)==1,maskk*npt2).astype('float32')
p2mask = np.tile(np.reshape(np.sum(np.reshape(segidx[tmpmaskidx],(maskk, npt2)),0)>0,(1,npt2)),(maskk,1))
p2mask = np.reshape(p2mask,maskk*npt2).astype('float32')
vmask = vmask*p2mask
tmppc1 = pc1[tmpmaskidx,:]
tmppc2 = np.tile(pc2,(maskk,1))
tmppc1 = tmppc1[np.logical_and(segidx[tmpmaskidx]>0,vmask),:]
tmppc2 = tmppc2[np.logical_and(segidx[tmpmaskidx]>0,vmask),:]
if tmppc1.shape[0]>5:
R, t = fit_motion(tmppc1, tmppc2)
else:
R = np.eye(3)
t = np.zeros(3)
maskdist = np.transpose(np.reshape(segidx[tmpmaskidx]>0.5,(maskk, npt2)),(1,0))
nrefine_iter = 3
for j in range(nrefine_iter):
curtrans = np.matmul(pc1, R)+t
tmpflag = np.zeros(npt2)
for k in range(maskk):
tmpflag = np.logical_or(tmpflag, maskdist[:,k] * (np.sqrt(np.sum((curtrans[maskidx[:,k],:]-pc2)**2,1))<eps) )
tmpmaskidx = maskidx[tmpflag,:]
tmpmaskdist = maskdist[tmpflag,:]
tmppc2 = pc2[tmpflag,:]
tmppc1 = pc1[tmpmaskidx[:,0],:]
tmpdist = np.inf*np.ones(tmppc2.shape[0])
for k in range(maskk):
dd = np.sqrt(np.sum((curtrans[tmpmaskidx[:,k],:]-tmppc2)**2,1))
flag = dd<tmpdist
flag = np.logical_and(flag, tmpmaskdist[:,k])
tmppc1[flag,:] = pc1[tmpmaskidx[flag,k],:]
tmpdist[flag] = dd[flag]
if tmppc1.shape[0]<3 or tmppc2.shape[0]<3:
R = np.eye(3)
t = np.zeros((1,3))
inlier_idx1 = np.zeros(npt1)==1
inlier_idx2 = np.zeros(npt2)==1
return R, t, inlier_idx1, inlier_idx2
else:
R, t = fit_motion(tmppc1, tmppc2)
inlier_idx2 = tmpflag
inlier_idx1 = np.zeros(npt1)==1
if tmppc1.shape[0]>0:
nbrs1 = NearestNeighbors(n_neighbors=1, algorithm='ball_tree').fit(pc1)
_, pc1flag = nbrs1.kneighbors(tmppc1)
pc1flag = np.unique(pc1flag)
inlier_idx1[pc1flag] = True
return R, t, inlier_idx1, inlier_idx2
def seg_merge(pc1, pcpred, segidx):
# pc1: N x 3
# pcpred: N x 3
# segidx: N
useg = np.unique(segidx)
nmode = len(useg)
npoint = pc1.shape[0]
if nmode==1:
return segidx
Rmodes = list()
tmodes = list()
for i in range(nmode):
R, t = fit_motion(pc1[segidx==useg[i],:], pcpred[segidx==useg[i],:])
Rmodes.append(R)
tmodes.append(t)
dres = np.zeros((npoint, nmode))
for i in range(nmode):
dres[:,i] = np.sqrt(np.sum((np.matmul(pc1, Rmodes[i])+tmodes[i]-pcpred)**2,1))
dmode = np.zeros((nmode, nmode))
for i in range(nmode):
dmode[i,:] = np.mean(dres[segidx==useg[i],:],0)
mindres = np.mean(np.diag(dmode))
dresth = np.minimum(0.06, mindres*3)
for i in range(2, nmode+1):
select_pool = list(itertools.combinations(np.arange(nmode), 2))
scores = np.zeros(len(select_pool))
for j in range(len(select_pool)):
scores[j] = np.mean(np.min(dmode[select_pool[j],:],0))
if np.min(scores)<dresth:
break
imx = np.argmin(scores,0)
imx = np.argmin(dmode[select_pool[imx],:],0)
segidx_merged = np.zeros_like(segidx)
for i in range(nmode):
segidx_merged[segidx==useg[i]]=imx[i]
if np.max(segidx_merged)==0 and nmode==2:
segidx_merged = segidx
return segidx_merged
def fps(pc1, Nout):
Nin = pc1.shape[0]
if Nout>Nin:
pcout1 = np.concatenate((pc1, pc1[np.random.choice(np.arange(Nin),Nout-Nin),:]),0)
else:
selectIdx = np.zeros(Nin)
seed = np.random.randint(Nin)
selectIdx[seed] = 1
count = 1
pd = np.sum((pc1-pc1[[seed],:])**2,1)
while count<Nout:
imx = np.argmax(pd)
selectIdx[imx] = 1
count += 1
pd = np.minimum(pd, np.sum((pc1-pc1[[imx],:])**2,1))
pcout1 = pc1[selectIdx==1,:]
return pcout1