-
Notifications
You must be signed in to change notification settings - Fork 25
/
loss.py
138 lines (101 loc) · 4.27 KB
/
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
# Copyright (c) 2018-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#
import numpy as np
import torch
def mpjpe(predicted, target):
"""
Mean per-joint position error (i.e. mean Euclidean distance),
often referred to as "Protocol #1" in many papers.
"""
assert predicted.shape == target.shape
return torch.mean(torch.norm(predicted - target, dim=len(target.shape) - 1))
def weighted_mpjpe(predicted, target, w):
"""
Weighted mean per-joint position error (i.e. mean Euclidean distance)
"""
assert predicted.shape == target.shape
assert w.shape[0] == predicted.shape[0]
return torch.mean(w * torch.norm(predicted - target, dim=len(target.shape) - 1))
def p_mpjpe(predicted, target):
"""
Pose error: MPJPE after rigid alignment (scale, rotation, and translation),
often referred to as "Protocol #2" in many papers.
"""
assert predicted.shape == target.shape
muX = np.mean(target, axis=1, keepdims=True)
muY = np.mean(predicted, axis=1, keepdims=True)
X0 = target - muX
Y0 = predicted - muY
normX = np.sqrt(np.sum(X0 ** 2, axis=(1, 2), keepdims=True))
normY = np.sqrt(np.sum(Y0 ** 2, axis=(1, 2), keepdims=True))
X0 /= normX
Y0 /= normY
H = np.matmul(X0.transpose(0, 2, 1), Y0)
U, s, Vt = np.linalg.svd(H)
V = Vt.transpose(0, 2, 1)
R = np.matmul(V, U.transpose(0, 2, 1))
# Avoid improper rotations (reflections), i.e. rotations with det(R) = -1
sign_detR = np.sign(np.expand_dims(np.linalg.det(R), axis=1))
V[:, :, -1] *= sign_detR
s[:, -1] *= sign_detR.flatten()
R = np.matmul(V, U.transpose(0, 2, 1)) # Rotation
tr = np.expand_dims(np.sum(s, axis=1, keepdims=True), axis=2)
a = tr * normX / normY # Scale
t = muX - a * np.matmul(muY, R) # Translation
# Perform rigid transformation on the input
predicted_aligned = a * np.matmul(predicted, R) + t
# Return MPJPE
return np.mean(np.linalg.norm(predicted_aligned - target, axis=len(target.shape) - 1))
def n_mpjpe(predicted, target):
"""
Normalized MPJPE (scale only), adapted from:
https://github.com/hrhodin/UnsupervisedGeometryAwareRepresentationLearning/blob/master/losses/poses.py
"""
assert predicted.shape == target.shape
norm_predicted = torch.mean(torch.sum(predicted ** 2, dim=3, keepdim=True), dim=2, keepdim=True)
norm_target = torch.mean(torch.sum(target * predicted, dim=3, keepdim=True), dim=2, keepdim=True)
scale = norm_target / norm_predicted
return mpjpe(scale * predicted, target)
def mean_velocity_error(predicted, target):
"""
Mean per-joint velocity error (i.e. mean Euclidean distance of the 1st derivative)
"""
assert predicted.shape == target.shape
velocity_predicted = np.diff(predicted, axis=0)
velocity_target = np.diff(target, axis=0)
return np.mean(np.linalg.norm(velocity_predicted - velocity_target, axis=len(target.shape) - 1))
def pose_align(predicted, target):
"""
Pose error: MPJPE after rigid alignment (scale, rotation, and translation),
often referred to as "Protocol #2" in many papers.
size: bx16x3
"""
assert predicted.shape == target.shape
muX = np.mean(target, axis=1, keepdims=True)
muY = np.mean(predicted, axis=1, keepdims=True)
X0 = target - muX
Y0 = predicted - muY
normX = np.sqrt(np.sum(X0 ** 2, axis=(1, 2), keepdims=True))
normY = np.sqrt(np.sum(Y0 ** 2, axis=(1, 2), keepdims=True))
X0 /= normX
Y0 /= normY
H = np.matmul(X0.transpose(0, 2, 1), Y0)
U, s, Vt = np.linalg.svd(H)
V = Vt.transpose(0, 2, 1)
R = np.matmul(V, U.transpose(0, 2, 1))
# Avoid improper rotations (reflections), i.e. rotations with det(R) = -1
sign_detR = np.sign(np.expand_dims(np.linalg.det(R), axis=1))
V[:, :, -1] *= sign_detR
s[:, -1] *= sign_detR.flatten()
R = np.matmul(V, U.transpose(0, 2, 1)) # Rotation
tr = np.expand_dims(np.sum(s, axis=1, keepdims=True), axis=2)
a = tr * normX / normY # Scale
t = muX - a * np.matmul(muY, R) # Translation
# Perform rigid transformation on the input
predicted_aligned = a * np.matmul(predicted, R) + t
# Return MPJPE
return predicted_aligned