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
| """
GazeAnywhere 模型架构
论文核心：transformer-based detector 融合冻结编码器特征
同时解决：主体定位 + 帧内外存在判断 + 视线目标热图估计
"""
import torch
import torch.nn as nn
from typing import Dict, Optional, Tuple
class GazeAnywhere(nn.Module):
"""
GazeAnywhere: 可提示的视线目标估计
输入：
- 图像 (B, 3, H, W)
- 文本提示 (B, L) 或 视觉提示 (B, 2) 归一化坐标
输出：
- 注视目标热图 (B, H//4, W//4)
- 帧内外存在 (B, 1)
- 主体定位 (B, 4) bbox
"""
def __init__(
self,
vision_model: str = 'CLIP-ViT-L/14',
text_model: str = 'CLIP-text',
embed_dim: int = 768,
num_queries: int = 1,
nhead: int = 8,
num_decoder_layers: int = 6
):
super().__init__()
self.vision_encoder = self._load_frozen_vision_encoder(vision_model)
self.vision_proj = nn.Linear(1024, embed_dim)
self.text_encoder = self._load_frozen_text_encoder(text_model)
self.text_proj = nn.Linear(512, embed_dim)
decoder_layer = nn.TransformerDecoderLayer(
d_model=embed_dim,
nhead=nhead,
dim_feedforward=embed_dim * 4,
batch_first=True,
dropout=0.1
)
self.decoder = nn.TransformerDecoder(
decoder_layer, num_layers=num_decoder_layers
)
self.queries = nn.Parameter(
torch.randn(num_queries, embed_dim) * 0.02
)
self.heatmap_head = nn.Sequential(
nn.Linear(embed_dim, embed_dim // 2),
nn.GELU(),
nn.Linear(embed_dim // 2, (224 // 4) ** 2),
nn.Sigmoid()
)
self.presence_head = nn.Sequential(
nn.Linear(embed_dim, embed_dim // 2),
nn.GELU(),
nn.Linear(embed_dim // 2, 1),
nn.Sigmoid()
)
self.bbox_head = nn.Sequential(
nn.Linear(embed_dim, embed_dim // 2),
nn.GELU(),
nn.Linear(embed_dim // 2, 4),
nn.Sigmoid()
)
def _load_frozen_vision_encoder(self, name: str):
"""加载冻结的视觉编码器"""
encoder = nn.TransformerEncoderLayer(d_model=1024, nhead=16)
for p in encoder.parameters():
p.requires_grad = False
return encoder
def _load_frozen_text_encoder(self, name: str):
"""加载冻结的文本编码器"""
encoder = nn.LSTM(512, 512, batch_first=True)
for p in encoder.parameters():
p.requires_grad = False
return encoder
def forward(
self,
image: torch.Tensor,
text_prompt: torch.Tensor = None,
visual_prompt: torch.Tensor = None
) -> Dict[str, torch.Tensor]:
"""
Args:
image: (B, 3, H, W)
text_prompt: (B, L, 512) 文本特征
visual_prompt: (B, 2) 归一化坐标 [x, y]
"""
B = image.shape[0]
visual_features = self.vision_proj(
self._extract_patches(image)
)
if text_prompt is not None:
prompt_embed = self.text_proj(text_prompt)
prompt_embed = prompt_embed.mean(dim=1, keepdim=True)
elif visual_prompt is not None:
prompt_embed = self._encode_visual_prompt(
visual_prompt, visual_features
)
else:
raise ValueError("需要文本或视觉提示")
queries = self.queries.unsqueeze(0).expand(B, -1, -1)
keys_values = torch.cat([visual_features, prompt_embed], dim=1)
decoded = self.decoder(queries, keys_values)
heatmap = self.heatmap_head(decoded)
heatmap = heatmap.view(B, -1, 56, 56)
presence = self.presence_head(decoded)
bbox = self.bbox_head(decoded)
return {
'heatmap': heatmap,
'presence': presence,
'bbox': bbox,
}
def _extract_patches(self, image: torch.Tensor) -> torch.Tensor:
"""提取图像 patch 特征（简化版）"""
B, C, H, W = image.shape
patch_size = 16
patches = image.view(
B, C, H // patch_size, patch_size, W // patch_size, patch_size
)
patches = patches.permute(0, 2, 4, 1, 3, 5).contiguous()
patches = patches.view(B, (H // patch_size) * (W // patch_size), -1)
return patches
def _encode_visual_prompt(
self, prompt: torch.Tensor, features: torch.Tensor
) -> torch.Tensor:
"""视觉提示编码：归一化坐标 → 特征选择"""
B = prompt.shape[0]
N = features.shape[1]
grid_size = int(N ** 0.5)
x, y = prompt[:, 0], prompt[:, 1]
idx = (y * grid_size + x).long()
selected = features[torch.arange(B), idx]
return selected.unsqueeze(1)
if __name__ == "__main__":
torch.manual_seed(42)
model = GazeAnywhere()
image = torch.randn(2, 3, 224, 224)
text = torch.randn(2, 10, 512)
visual = torch.tensor([[0.5, 0.3], [0.2, 0.7]])
out_text = model(image, text_prompt=text)
print("文本提示输出:")
print(f" 热图: {out_text['heatmap'].shape}")
print(f" 存在: {out_text['presence'].shape}")
print(f" 定位: {out_text['bbox'].shape}")
out_vis = model(image, visual_prompt=visual)
print("\n视觉提示输出:")
print(f" 热图: {out_vis['heatmap'].shape}")
print(f" 存在: {out_vis['presence'].shape}")
print(f" 定位: {out_vis['bbox'].shape}")
|
