-
Notifications
You must be signed in to change notification settings - Fork 6
/
epc_loss.py
215 lines (159 loc) · 9.05 KB
/
epc_loss.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
import torch
import torch.nn as nn
from sklearn.cluster import DBSCAN
from utils_rigid import kabsch_transformation_estimation
import numpy as np
import copy
from pointnet2 import pointnet2_utils
EPS = 0.6 # used in DBSCAN for clustering
# 0.6 - waymo 0.4 - kitti 0.3 - lyft
L_K = 6 # num of neighboring points to calculate Laplacian Coordinates
class ChamferLoss(nn.Module):
def __init__(self):
super(ChamferLoss, self).__init__()
self.point_criterion = nn.L1Loss(reduction='mean')
def forward(self, x1, y1):
'''
n1 = n2 = 8192
:param x1: (1, 3, 8192) -- (1, 8192, 3) -- (1, 1, 8192, 3)
:param y1: (1, 3, 8192) -- (1, 8192, 3) -- (1, 8192, 1, 3)
:return: loss
'''
x = torch.transpose(x1, 1, 2)
y = torch.transpose(y1, 1, 2)
x = x.unsqueeze(1)
y = y.unsqueeze(2)
dist = torch.sqrt(1e-8 + torch.sum(torch.pow(x - y, 2), 3)) # bs, ny, nx --- pointwise dist
min1, _ = torch.min(dist, 1)
min2, _ = torch.min(dist, 2)
return min1.mean() + min2.mean()
class DistanceLoss(nn.Module):
def __init__(self):
super(DistanceLoss, self).__init__()
def forward(self, x1, y1):
dist = torch.sqrt(1e-6 + torch.sum(torch.pow(x1 - y1, 2), 1))
return dist.mean()
class L1Loss(nn.Module):
def __init__(self):
super(L1Loss, self).__init__()
self.point_criterion = nn.L1Loss(reduction='mean')
def forward(self, x1, y1):
return self.point_criterion(x1, y1)
def RigidReconstruction(pc, sf, min_samples, metric, eps, min_p_cluster):
cluster_estimator = DBSCAN(min_samples=min_samples, metric=metric, eps=eps)
labels_x = cluster_estimator.fit_predict(pc.transpose(1, 2).squeeze(0).cpu().numpy())
clusters_x = []
for class_label in np.unique(labels_x):
if class_label != -1 and np.where(labels_x == class_label)[0].shape[0] >= min_p_cluster:
clusters_x.append(np.where(labels_x == class_label)[0])
pc_rigid = (pc + sf).transpose(1, 2).squeeze(0) # (1, 8192, 3)
for cluster_i in clusters_x:
cluster_pc = pc[:, :, cluster_i].transpose(1, 2) # pc
cluster_sf = sf[:, :, cluster_i].transpose(1, 2) # sf
# cluster_recon = (cluster_pc + cluster_sf).squeeze(0)
cluster_recon = (cluster_pc + cluster_sf)
R_cluster, t_cluster, _, _ = kabsch_transformation_estimation(cluster_pc, cluster_recon) # estimate R and T
rigid_recon = (torch.matmul(R_cluster, cluster_pc.transpose(1, 2)) + t_cluster).detach().squeeze(0).transpose(0,1)
pc_rigid[cluster_i, :] = rigid_recon
return pc_rigid, clusters_x
class RigidLoss(nn.Module):
def __init__(self, device, eps=EPS, min_samples=5, metric='euclidean', min_p_cluster=30):
super(RigidLoss, self).__init__()
self.device = device
self.min_p_cluster = min_p_cluster
self.cluster_estimator = DBSCAN(min_samples=min_samples, metric=metric, eps=eps)
self.rigidity_criterion = nn.L1Loss(reduction='mean')
#def forward(self, x, y, x0, y0, sf_x):
def forward(self, pc ,sf, pc_t, sf_t):
labels_x = self.cluster_estimator.fit_predict(pc.transpose(1, 2).squeeze(0).cpu().numpy())
clusters_x = []
for class_label in np.unique(labels_x):
if class_label != -1 and np.where(labels_x == class_label)[0].shape[0] >= self.min_p_cluster:
clusters_x.append(np.where(labels_x == class_label)[0])
rigid_loss = torch.tensor(0.0).cuda().to(self.device)
pc_rigid = (pc_t + sf_t).transpose(1, 2).squeeze(0) # (1, 8196, 3)
for cluster_i in clusters_x:
cluster_pc = pc[:, :, cluster_i].transpose(1, 2) # pc -- student input
cluster_sf = sf[:, :, cluster_i].transpose(1, 2) # sf -- student pred
cluster_pc_t = pc_t[:, :, cluster_i].transpose(1, 2) # pc_t -- teacher input
cluster_sf_t = sf_t[:, :, cluster_i].transpose(1, 2) # sf_t -- teacher pred
cluster_recon = (cluster_pc + cluster_sf).squeeze(0)
cluster_recon_t = (cluster_pc_t + cluster_sf_t).squeeze(0)
R_cluster, t_cluster, _, _ = kabsch_transformation_estimation(cluster_pc_t, cluster_recon_t) # estimate R and T
rigid_recon_t = (torch.matmul(R_cluster, cluster_pc_t.transpose(1, 2)) + t_cluster).detach().squeeze(0).transpose(0,1)
pc_rigid[cluster_i, :] = rigid_recon_t
# rigid_loss += self.rigidity_criterion(cluster_recon, rigid_recon_t)
# cluster_x_i, (cluster_x_i + sf_x_i)
#pc_rigid = RigidReconstruction(pc_t, sf_t)
rigid_loss = self.rigidity_criterion((pc + sf).transpose(1, 2).squeeze(0), pc_rigid)
return rigid_loss
class ConsistLoss(nn.Module):
def __init__(self):
super(ConsistLoss, self).__init__()
self.point_criterion = nn.L1Loss(reduction='mean')
self.idx = 0
def forward(self, input_t, sf_t, y1, pred):
'''
:param input_t: pc1_target pc1_te
:param sf_t: teacher_predicted_sf sf_te
:param y1: pc2_target pc2_te
:param pred: student_prediction (pc1_target_2 + student_predicted_sf) pred_st
:return: loss
'''
# DR -- reconstructs rigid bodies
rigid_recon, clusters = RigidReconstruction(input_t, sf_t, min_samples=5, metric='euclidean', eps=EPS, min_p_cluster=30)
self.idx += 1
# x1 = input_t + sf_t # teacher_prediction (pc1_target + teacher_predicted_sf)
x1 = torch.transpose(rigid_recon, 0, 1).unsqueeze(0)
# x -- pc1+sf (1, 3, 8192) -- (1, 8192, 3) -- (1, 1, 8192, 3)
# y -- pc2 (1, 3, 8192) -- (1, 8192, 3) -- (1, 8192, 1, 3)
x = torch.transpose(x1, 1, 2)
y = torch.transpose(y1, 1, 2)
x = x.unsqueeze(1)
y = y.unsqueeze(2)
# Correspondence Refinement
squared_dist = torch.sum(torch.pow(x - y, 2), 3) # bs, ny, nx
_, k_idx = torch.topk(squared_dist, L_K, dim=1, largest=False, sorted=False)
k_idx = k_idx.permute(0, 2, 1).contiguous() # (1, 8192, 10)
grouped_y = pointnet2_utils.grouping_operation(y1, k_idx.int()).permute(0, 2, 3, 1)
laplace_y = torch.sum(grouped_y - x1.permute(0, 2, 1).unsqueeze(2), dim=2) / float(L_K - 1)
squared_dist_self = torch.sum(torch.pow(x - x.permute(0, 2, 1, 3), 2), 3)
_, k_idx_self = torch.topk(squared_dist_self, L_K, dim=1, largest=False, sorted=False)
k_idx_self = k_idx_self.permute(0, 2, 1).contiguous() # (1, 8192, 10)
grouped_x = pointnet2_utils.grouping_operation(x1, k_idx_self.int()).permute(0, 2, 3, 1)
laplace_x = torch.sum(grouped_x - x1.permute(0, 2, 1).unsqueeze(2), dim=2) / float(L_K - 1)
laplace_reduced = laplace_x - laplace_y
rigid_refine = copy.deepcopy(rigid_recon)
for cluster_i in clusters:
rigid_refine[cluster_i, :] = rigid_recon[cluster_i, :] - torch.mean(laplace_reduced[:, cluster_i, :],
dim=1).squeeze(0)
loss = self.point_criterion(rigid_refine.unsqueeze(0), pred.permute(0, 2, 1))
return loss
def pred_refine(input_t, sf_t, y1):
# teacher_input_pc1, teacher_output, input_pc2
x1 = input_t + sf_t
x = torch.transpose(x1, 1, 2)
y = torch.transpose(y1, 1, 2) # x -- pc1+sf y -- pc2
x = x.unsqueeze(1)
y = y.unsqueeze(2)
squared_dist = torch.sum(torch.pow(x - y, 2), 3) # bs, ny, nx
_, k_idx = torch.topk(squared_dist, L_K, dim=1, largest=False, sorted=False)
k_idx = k_idx.permute(0, 2, 1).contiguous() # (1, 8192, 10)
grouped_y = pointnet2_utils.grouping_operation(y1, k_idx.int()).permute(0, 2, 3, 1)
laplace_y = torch.sum(grouped_y - x1.permute(0, 2, 1).unsqueeze(2), dim=2) / float(L_K - 1)
squared_dist_self = torch.sum(torch.pow(x - x.permute(0, 2, 1, 3), 2), 3)
_, k_idx_self = torch.topk(squared_dist_self, L_K, dim=1, largest=False, sorted=False)
k_idx_self = k_idx_self.permute(0, 2, 1).contiguous() # (1, 8192, 10)
grouped_x = pointnet2_utils.grouping_operation(x1, k_idx_self.int()).permute(0, 2, 3, 1)
laplace_x = torch.sum(grouped_x - x1.permute(0, 2, 1).unsqueeze(2), dim=2) / float(L_K - 1)
laplace_reduced = laplace_x - laplace_y
laplace_dist = torch.sum(torch.pow(laplace_reduced, 2), 2)
x_zeros = torch.zeros_like(x1)
laplace_dist = laplace_dist.unsqueeze(2).repeat(1, 1, 3)
#x_confident = torch.where(laplace_dist < 0.2, x1.permute(0, 2, 1) - laplace_reduced, pred.permute(0, 2, 1))
rigid_recon, clusters = RigidReconstruction(input_t, sf_t, min_samples=5, metric='euclidean', eps=EPS, min_p_cluster=30)
for cluster_i in clusters:
rigid_recon[cluster_i, :] = rigid_recon[cluster_i, :] - torch.mean(laplace_reduced[:, cluster_i, :], dim=1).squeeze(0)
return rigid_recon
if __name__ == '__main__':
pass