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
| import torch
import torch.nn as nn
class StudentPoseFormer(nn.Module):
"""
StudentPoseFormer: 从mmWave点云到3D人体网格
三大创新：
1. SAM 3D Body教师 — 零训练全身体标注
2. Set编码 + masked attention pooling
3. CVAE多假设头 — 输出姿态均值+逐关节方差
"""
def __init__(self,
n_joints: int = 70,
n_mesh_vertices: int = 18439,
d_model: int = 256,
n_heads: int = 8,
n_temporal_layers: int = 4,
n_accumulated_frames: int = 3):
super().__init__()
self.n_joints = n_joints
self.n_vertices = n_mesh_vertices
self.n_accumulated = n_accumulated_frames
self.point_encoder = nn.Sequential(
nn.Linear(5, 128),
nn.GELU(),
nn.Linear(128, d_model)
)
self.masked_attn_pool = nn.MultiheadAttention(
embed_dim=d_model,
num_heads=n_heads,
batch_first=True
)
encoder_layer = nn.TransformerEncoderLayer(
d_model=d_model,
nhead=n_heads,
dim_feedforward=d_model * 4,
dropout=0.1,
batch_first=True,
activation='gelu'
)
self.temporal_transformer = nn.TransformerEncoder(
encoder_layer, n_temporal_layers
)
self.cvae_encoder = nn.Sequential(
nn.Linear(d_model, d_model),
nn.GELU(),
nn.Linear(d_model, d_model * 2)
)
self.joint_decoder = nn.Sequential(
nn.Linear(d_model, d_model),
nn.GELU(),
nn.Linear(d_model, n_joints * 3)
)
self.uncertainty_decoder = nn.Sequential(
nn.Linear(d_model, d_model),
nn.GELU(),
nn.Linear(d_model, n_joints)
)
self.mesh_regressor = nn.Linear(n_joints * 3, n_mesh_vertices * 3)
def forward(self, point_clouds: list, mask: torch.Tensor = None):
"""
Args:
point_clouds: list of (N_i, 5) tensors, 每帧N_i个点
mask: (B, T, max_N) bool mask for padding
Returns:
joints: (B, T, 70, 3) 关节位置
uncertainties: (B, T, 70) 逐关节不确定性
mesh_vertices: (B, T, 18439, 3) 网格顶点
"""
batch_size = len(point_clouds)
T = len(point_clouds[0]) if isinstance(point_clouds[0], list) else 1
frame_features = []
for b in range(batch_size):
for t in range(T):
points = point_clouds[b][t] if T > 1 else point_clouds[b]
point_feat = self.point_encoder(points)
feat = point_feat.unsqueeze(0)
query = torch.mean(feat, dim=1, keepdim=True)
pooled, _ = self.masked_attn_pool(query, feat, feat)
frame_features.append(pooled.squeeze(0))
x = torch.stack(frame_features).view(batch_size, T, -1)
x = self.temporal_transformer(x)
cvae_out = self.cvae_encoder(x)
mean, log_var = cvae_out.chunk(2, dim=-1)
z = mean + torch.exp(0.5 * log_var) * torch.randn_like(mean)
joints = self.joint_decoder(z).view(batch_size, T, self.n_joints, 3)
uncertainties = torch.exp(self.uncertainty_decoder(z))
mesh = self.mesh_regressor(
joints.view(batch_size, T, -1)
).view(batch_size, T, self.n_vertices, 3)
return {
'joints': joints,
'uncertainties': uncertainties,
'mesh_vertices': mesh,
'kl_mean': mean,
'kl_log_var': log_var
}
if __name__ == "__main__":
model = StudentPoseFormer()
np.random.seed(42)
point_clouds = [
[torch.randn(np.random.randint(4, 10), 5) for _ in range(3)]
for _ in range(2)
]
output = model(point_clouds)
print(f"输入: batch=2, 3帧累积")
print(f"关节输出: {output['joints'].shape}")
print(f"不确定性: {output['uncertainties'].shape}")
print(f"网格顶点: {output['mesh_vertices'].shape}")
print(f"\n不确定性最高的关节（雷达看不到的地方）:")
unc = output['uncertainties'][0, 0].detach().numpy()
joint_names = ["hip_r", "hip_l", "spine", "neck", "head",
"shoulder_r", "elbow_r", "wrist_r",
"shoulder_l", "elbow_l", "wrist_l",
"knee_r", "ankle_r", "knee_l", "ankle_l"]
for name, u in sorted(zip(joint_names[:15], unc[:15]), key=lambda x: -x[1]):
bar = '█' * int(u * 10)
print(f" {name:<15} {u:.4f} {bar}")
|
