-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathcorrection_skeleton.py
135 lines (105 loc) · 6.02 KB
/
correction_skeleton.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
import torch
import torch.nn as nn
import numpy as np
from pytorch3d.transforms import matrix_to_rotation_6d, quaternion_to_matrix, rotation_6d_to_matrix, matrix_to_quaternion
from model.layers import ST_GCNN_layer
class ObjProjector(nn.Module):
def __init__(self, args):
super(ObjProjector, self).__init__()
self.args = args
num_channels = args.embedding_dim
self.n_pre = 20
self.st_gcnns_relative=nn.ModuleList()
self.st_gcnns_relative.append(ST_GCNN_layer(9,32,[1,1],1,self.n_pre,
args.num_joints,args.dropout,version=0))
self.st_gcnns_relative.append(ST_GCNN_layer(32,16,[1,1],1,self.n_pre,
args.num_joints,args.dropout,version=0))
self.st_gcnns_relative.append(ST_GCNN_layer(16,32,[1,1],1,self.n_pre,
args.num_joints,args.dropout,version=0))
self.st_gcnns_relative.append(ST_GCNN_layer(32,9,[1,1],1,self.n_pre,
args.num_joints,args.dropout,version=0))
self.st_gcnns=nn.ModuleList()
self.st_gcnns.append(ST_GCNN_layer(9,32,[1,1],1,self.n_pre,
1,args.dropout,version=0))
self.st_gcnns.append(ST_GCNN_layer(32,16,[1,1],1,self.n_pre,
1,args.dropout,version=0))
self.st_gcnns.append(ST_GCNN_layer(16,32,[1,1],1,self.n_pre,
1,args.dropout,version=0))
self.st_gcnns.append(ST_GCNN_layer(32,9,[1,1],1,self.n_pre,
1,args.dropout,version=0))
self.st_gcnns_all=nn.ModuleList()
self.st_gcnns_all.append(ST_GCNN_layer(9,64,[1,1],1,self.n_pre,
args.num_joints+1,args.dropout,version=2))
self.st_gcnns_all.append(ST_GCNN_layer(64,32,[1,1],1,self.n_pre,
args.num_joints+1,args.dropout,version=2))
self.st_gcnns_all.append(ST_GCNN_layer(32,64,[1,1],1,self.n_pre,
args.num_joints+1,args.dropout,version=2))
self.st_gcnns_all.append(ST_GCNN_layer(64,9,[1,1],1,self.n_pre,
args.num_joints+1,args.dropout,version=2))
self.dct_m, self.idct_m = self.get_dct_matrix(args.past_len + args.future_len)
def get_dct_matrix(self, N, is_torch=True):
dct_m = np.eye(N)
for k in np.arange(N):
for i in np.arange(N):
w = np.sqrt(2 / N)
if k == 0:
w = np.sqrt(1 / N)
dct_m[k, i] = w * np.cos(np.pi * (i + 1 / 2) * k / N)
idct_m = np.linalg.inv(dct_m)
if is_torch:
dct_m = torch.from_numpy(dct_m)
idct_m = torch.from_numpy(idct_m)
return dct_m, idct_m
def forward(self, obj_angles, obj_trans, human_points):
# NOTE: align data format
# obj_angles: T,B,4
# obj_trans: T,B,3
# human_points: T,B,N_joints,3
obj_angles_gt = obj_angles.clone()
quat_correct = torch.cat([obj_angles[:,:,-1,None], obj_angles[:,:,-4:-1]],dim=2)
obj_angles = matrix_to_rotation_6d(quaternion_to_matrix(quat_correct))
assert not obj_angles.isnan().any()
obj_trans_gt = obj_trans.clone()
obj_angles_p, obj_trans_p = self.sample(obj_angles, obj_trans, human_points)
return obj_angles_p, obj_trans_p, obj_angles_gt, obj_trans_gt
def sample(self, obj_angles, obj_trans, human_points):
# TODO: align data format
# obj_angles: T,B,4
# obj_trans: T,B,3
# human_points: T,B,N_joints,3
quat_correct = torch.cat([obj_angles[:,:,-1,None], obj_angles[:,:,-4:-1]],dim=2)
obj_angles = matrix_to_rotation_6d(quaternion_to_matrix(quat_correct))
dct_m = self.dct_m.to(obj_angles.device).float()
idct_m = self.idct_m.to(obj_angles.device).float()
idx_pad = list(range(self.args.past_len)) + [self.args.past_len - 1] * self.args.future_len
obj_trans_relative = obj_trans.unsqueeze(2) - human_points
obj_relative = torch.cat([obj_angles.unsqueeze(2).repeat(1, 1, obj_trans_relative.shape[2], 1), obj_trans_relative], dim=3)[idx_pad]
T, B, P, C = obj_relative.shape
obj_relative = obj_relative.permute(1, 0, 3, 2).contiguous().view(B, T, C * P)
obj_relative = torch.matmul(dct_m[:self.n_pre], obj_relative).view(B, -1, C, P).permute(0, 2, 1, 3).contiguous() # B C T P
x = obj_relative.clone()
for gcn in (self.st_gcnns_relative):
x = gcn(x)
obj_relative = obj_relative + x
human_trans = human_points.permute(1, 0, 3, 2).contiguous().view(B, T, -1)
human_trans = torch.matmul(dct_m[:self.n_pre], human_trans).view(B, -1, 3, P).permute(0, 2, 1, 3).contiguous() # B C T P
obj_multi = torch.cat([obj_relative[:, :6, :, :], obj_relative[:, 6:9, :, :] + human_trans], dim=1)
obj_gt = torch.cat([obj_angles, obj_trans], dim=2)
obj = obj_gt[idx_pad].unsqueeze(2)
obj = obj.permute(1, 0, 3, 2).contiguous().view(B, T, C * 1)
obj = torch.matmul(dct_m[:self.n_pre], obj).view(B, -1, C, 1).permute(0, 2, 1, 3).contiguous() # B C T P
x = obj.clone()
for gcn in (self.st_gcnns):
x = gcn(x)
obj = obj + x
obj = torch.cat([obj, obj_multi], dim=3)
x = obj.clone()
for gcn in (self.st_gcnns_all):
x = gcn(x)
obj = obj + x
obj = obj.permute(0, 2, 1, 3).contiguous().view(B, -1, C * (P+1))
results = torch.matmul(idct_m[:, :self.n_pre], obj).view(B, T, C, P+1).permute(1, 0, 3, 2)[:, :, 0, :9]
obj_angles_p = matrix_to_quaternion(rotation_6d_to_matrix(results[:,:,:6]))
obj_angles_p = torch.cat([obj_angles_p[:,:,1:4], obj_angles_p[:,:,0,None]],dim=2)
obj_trans_p = results[:,:,6:9]
return obj_angles_p, obj_trans_p