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
| import torch
import torch.nn as nn
import numpy as np
import cv2
class DepthFromDefocus(nn.Module):
"""散焦深度估计: 利用模糊程度推断深度"""
def __init__(self):
super().__init__()
self.blur_estimator = nn.Sequential(
nn.Conv2d(1, 32, 3, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, 3, padding=1),
nn.ReLU(),
nn.Conv2d(64, 1, 1),
nn.Sigmoid()
)
def estimate_depth(self, image):
"""
利用散焦模糊估计深度
原理: 近距离 → 大光圈模糊, 远距离 → 小光圈模糊
"""
gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
gray_t = torch.from_numpy(gray).float().unsqueeze(0).unsqueeze(0) / 255.0
blur_map = self.blur_estimator(gray_t)
depth = 1.0 - blur_map
return depth
def extract_head_pose(self, image):
"""从散焦深度估计头部姿态"""
depth = self.estimate_depth(image)
dx = torch.gradient(depth.squeeze(), dim=-1)[0]
dy = torch.gradient(depth.squeeze(), dim=-2)[0]
pitch = torch.atan2(dy.mean(), 1.0).item()
yaw = torch.atan2(dx.mean(), 1.0).item()
roll = 0.0
tx, ty, tz = 0.0, 0.0, depth.mean().item()
return np.array([pitch, yaw, roll, tx, ty, tz])
class IrisLocalizer(nn.Module):
"""虹膜定位器"""
def __init__(self):
super().__init__()
self.detector = nn.Sequential(
nn.Conv2d(1, 32, 3, padding=1), nn.ReLU(),
nn.Conv2d(32, 64, 3, stride=2, padding=1), nn.ReLU(),
nn.Conv2d(64, 128, 3, stride=2, padding=1), nn.ReLU(),
nn.Conv2d(128, 2, 1)
)
def forward(self, eye_region):
"""定位虹膜中心"""
return self.detector(eye_region)
class MarkerFreeGazeEstimator(nn.Module):
"""完整无标记视线估计"""
def __init__(self, feature_dim=8, num_calibration_points=9):
super().__init__()
self.depth_module = DepthFromDefocus()
self.iris_module = IrisLocalizer()
self.mapping = nn.Linear(feature_dim, 2)
self.num_cal = num_calibration_points
def forward(self, face_image, eye_region):
"""
Args:
face_image: (H, W, 3) 面部图像
eye_region: (1, 1, 64, 64) 眼部裁剪
Returns:
gaze_point: (2,) 注视点 (x, y)
"""
head_pose = self.depth_module.extract_head_pose(face_image)
iris_center = self.iris_module(eye_region)
iris_x, iris_y = iris_center.squeeze().tolist()
features = torch.tensor([
head_pose[0], head_pose[1], head_pose[2],
head_pose[3], head_pose[4], head_pose[5],
iris_x, iris_y
])
gaze_point = self.mapping(features)
return gaze_point, features
if __name__ == "__main__":
model = MarkerFreeGazeEstimator()
print("Marker-free Gaze Estimation:")
print(f" Features: 8D (6 head pose + 2 iris)")
print(f" Calibration points: {model.num_cal}")
print(f" No external markers needed")
print(f" Single 2D camera only")
|
