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
| import torch
import torch.nn as nn
import torch.nn.functional as F
class RotationConstrainedGazeFusion(nn.Module):
"""
旋转约束跨视角视线融合
WACV 2024 论文核心：
多摄像头视角不同，需旋转约束对齐
关键：
1. 旋转不变特征提取
2. 跨视角特征融合
3. 约束优化视线估计
"""
def __init__(self,
num_views: int = 2,
feature_dim: int = 256,
gaze_dim: int = 2):
super().__init__()
self.view_encoder = nn.Sequential(
nn.Conv2d(3, 64, 7, 2, 3),
nn.ReLU(inplace=True),
nn.Conv2d(64, 128, 3, 2, 1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 3, 2, 1),
nn.ReLU(inplace=True),
nn.AdaptiveAvgPool2d((1, 1))
)
self.rotation_constraint = RotationConstraintModule(feature_dim)
self.fusion_module = CrossViewFusion(num_views, feature_dim)
self.gaze_head = nn.Linear(feature_dim, gaze_dim)
self.num_views = num_views
def forward(self,
multi_view_images: torch.Tensor,
view_rotations: torch.Tensor) -> dict:
"""
前向传播
Args:
multi_view_images: 多视角图像, shape=(B, num_views, C, H, W)
view_rotations: 各视角相对旋转角度, shape=(B, num_views, 3) (pitch, yaw, roll)
Returns:
dict: {
"fused_gaze": 融合视线估计, shape=(B, 2)
"view_gazes": 各视角视线估计, shape=(B, num_views, 2)
"fusion_weights": 融合权重, shape=(B, num_views)
"rotation_error": 旋转约束误差, shape=(B, 1)
}
"""
B = multi_view_images.size(0)
view_features = []
for i in range(self.num_views):
view_img = multi_view_images[:, i]
features = self.view_encoder(view_img).view(B, -1)
view_features.append(features)
view_features_stack = torch.stack(view_features, dim=1)
aligned_features, rotation_error = self.rotation_constraint(
view_features_stack, view_rotations
)
fused_features, fusion_weights = self.fusion_module(aligned_features)
fused_gaze = self.gaze_head(fused_features)
view_gazes = []
for i in range(self.num_views):
view_gaze = self.gaze_head(aligned_features[:, i])
view_gazes.append(view_gaze)
view_gazes = torch.stack(view_gazes, dim=1)
return {
"fused_gaze": fused_gaze,
"view_gazes": view_gazes,
"fusion_weights": fusion_weights,
"rotation_error": rotation_error
}
class RotationConstraintModule(nn.Module):
"""
旋转约束模块
论文方法：
不同视角需旋转对齐才能融合
约束：
旋转后的视线应满足几何一致性
"""
def __init__(self, feature_dim: int):
super().__init__()
self.rotation_encoder = nn.Sequential(
nn.Linear(3, 64),
nn.ReLU(inplace=True),
nn.Linear(64, feature_dim)
)
self.alignment_head = nn.Linear(feature_dim * 2, feature_dim)
def forward(self,
view_features: torch.Tensor,
rotations: torch.Tensor) -> tuple:
"""
旋转对齐
Args:
view_features: 各视角特征, shape=(B, num_views, feature_dim)
rotations: 旋转角度, shape=(B, num_views, 3)
Returns:
aligned_features: 对齐后的特征
rotation_error: 旋转一致性误差
"""
B, num_views, feature_dim = view_features.shape
rotation_features = self.rotation_encoder(rotations)
aligned_features = []
for i in range(num_views):
combined = torch.cat([
view_features[:, i],
rotation_features[:, i]
], dim=1)
aligned = self.alignment_head(combined)
aligned_features.append(aligned)
aligned_features = torch.stack(aligned_features, dim=1)
rotation_error = self._compute_rotation_consistency(
aligned_features, rotations
)
return aligned_features, rotation_error
def _compute_rotation_consistency(self,
features: torch.Tensor,
rotations: torch.Tensor) -> torch.Tensor:
"""
旋转一致性计算
不同视角估计的视线，旋转到同一坐标系后应一致
"""
similarity_matrix = F.cosine_similarity(
features[:, 0].unsqueeze(1),
features[:, 1].unsqueeze(1),
dim=2
)
consistency_error = 1 - similarity_matrix.mean()
return consistency_error.unsqueeze(1)
class CrossViewFusion(nn.Module):
"""
跨视角融合模块
论文方法：
加权融合多视角视线估计
权重基于：
- 视角质量（遮挡、光照）
- 旋转角度（正面权重高）
"""
def __init__(self, num_views: int, feature_dim: int):
super().__init__()
self.weight_predictor = nn.Sequential(
nn.Linear(feature_dim * num_views, 128),
nn.ReLU(inplace=True),
nn.Linear(128, num_views)
)
def forward(self, aligned_features: torch.Tensor) -> tuple:
"""
融合
Args:
aligned_features: 对齐后的特征, shape=(B, num_views, feature_dim)
Returns:
fused_features: 融合特征
fusion_weights: 融合权重
"""
B, num_views, feature_dim = aligned_features.shape
features_flat = aligned_features.view(B, -1)
weights_logits = self.weight_predictor(features_flat)
fusion_weights = F.softmax(weights_logits, dim=1)
fused_features = torch.sum(
aligned_features * fusion_weights.unsqueeze(2),
dim=1
)
return fused_features, fusion_weights
if __name__ == "__main__":
model = RotationConstrainedGazeFusion(num_views=2)
multi_view_images = torch.randn(2, 2, 3, 224, 224)
view_rotations = torch.tensor([
[[0, 0, 0], [0, 30, 0]],
[[0, 15, 0], [0, 45, 0]],
], dtype=torch.float32)
output = model(multi_view_images, view_rotations)
print("=== 多视角视线融合结果 ===")
print(f"融合视线: {output['fused_gaze']}")
print(f"各视角视线: {output['view_gazes']}")
print(f"融合权重: {output['fusion_weights']}")
print(f"旋转一致性误差: {output['rotation_error']}")
|