哪位老兄帮忙哈，脸部特征点提取的问题

day270010678 · 发表于昨天 11:01

星级打分

1
2
3
4
5

平均分:NAN 参与人数:0 我的评分:未评

彻底换了deepfacelab框架，基于pytorch框架构建的，丫的光这个面部识别参数调整了一个小时了，愣是不能完美。哪位老兄帮忙哈，给个逻辑，或者重构哈。

def get_pts_from_predict(self, a, center, scale):
"""
Extract points from the model prediction heatmaps
"""
a_ch, a_h, a_w = a.shape
print(f"get_pts_from_predict: Input shape = ({a_ch}, {a_h}, {a_w}), dtype = {a.dtype}")
# Make sure we have the right number of channels (should be 68 for landmarks)
if a_ch != 68:
print(f"Warning: Expected 68 channels for landmarks, got {a_ch}. Using fallback landmarks.")
base_landmarks = LandmarksProcessor.landmarks_2D
# Create fallback landmarks based on center
fallback_pts = base_landmarks.astype(np.float32) * 100 + np.array([center[0], center[1]], dtype=np.float32)
return fallback_pts
b = a.reshape((a_ch, a_h*a_w))
c = b.argmax(1).reshape((a_ch, 1)).repeat(2, axis=1).astype(np.float32) # Force to float32
c[:,0] %= a_w
c[:,1] = np.apply_along_axis(lambda x: np.floor(x / a_w), 0, c[:,1])
for i in range(a_ch):
pX, pY = int(c[i,0]), int(c[i,1])
# Use dynamic boundary check instead of hardcoding 63
if pX > 0 and pX < a_w-1 and pY > 0 and pY < a_h-1:
diff = np.array([a[i,pY,pX+1]-a[i,pY,pX-1], a[i,pY+1,pX]-a[i,pY-1,pX]], dtype=np.float32) # Force to float32
c[i] += np.sign(diff)*0.25
c += 0.5
# Apply a scaling factor to adjust landmark size if needed
pts = np.array([self.transform(c[i], center, scale, a_w) for i in range(a_ch)], dtype=np.float32) # Ensure output is float32
# Fine-tune adjustment based on face size
# Calculate the actual face width and height from the bounding box
face_width = scale * 195.0 # This is the actual face width in pixels
# Use a reference face size (e.g., 200 pixels) as baseline
reference_face_size = 200.0
# Calculate adjustment factor based on actual face size
adjustment_factor = face_width / reference_face_size
# Apply a small adjustment to reduce the spread slightly
# Based on historical experience, use a factor between 0.95 and 0.98
fine_tune_factor = 0.97 # Reduce spread by 3%
# Apply adjustment to landmarks
if adjustment_factor > 0:
# Calculate centroid of landmarks
centroid = np.mean(pts, axis=0)
# Scale distances from centroid by the adjustment factor
pts = centroid + (pts - centroid) * adjustment_factor * fine_tune_factor
return pts

复制代码

day270010678 · 发表于昨天 11:02

精细化了我发现还是有问题
# Apply a small adjustment to reduce the spread slightly
# Based on historical experience, use a factor between 0.95 and 0.98
fine_tune_factor = 0.97 # Reduce spread by 3%

day270010678 · 发表于昨天 11:26

def transform(self, point, center, scale, resolution):
"""
Transform a point from model space to image space
"""
pt = np.array([point[0], point[1], 1.0], dtype=np.float32) # Ensure float32
h = 200.0 * scale
m = np.eye(3, dtype=np.float32) # Ensure float32
m[0,0] = resolution / h
m[1,1] = resolution / h
m[0,2] = resolution * ( -center[0] / h + 0.5 )
m[1,2] = resolution * ( -center[1] / h + 0.5 )
m = np.linalg.inv(m)
# Apply a small adjustment to the transformed points based on actual face size
# This is based on empirical observation that PyTorch version tends to produce slightly smaller landmarks
adjusted_pt = np.matmul(m, pt)[0:2].astype(np.float32)
# Calculate the actual face width and height from the bounding box
face_width = scale * 195.0 # This is the actual face width in pixels
# Use a reference face size (e.g., 200 pixels) as baseline
reference_face_size = 200.0
# Calculate adjustment factor based on actual face size
adjustment_factor = face_width / reference_face_size
# Apply a small adjustment to reduce the spread slightly
fine_tune_factor = 0.98 # Reduce by 2%
# Apply adjustment to the transformed point
if adjustment_factor > 0:
# Calculate centroid of landmarks (approximate)
centroid = np.array([center[0], center[1]], dtype=np.float32)
# Scale distances from centroid by the adjustment factor
adjusted_pt = centroid + (adjusted_pt - centroid) * adjustment_factor * fine_tune_factor
return adjusted_pt

复制代码

在transform方法中添加一个基于实际人脸尺寸的动态调整机制也不行，草了。

day270010678 · 发表于昨天 11:29

基于热图峰值位置的精确调整，也不行

def get_pts_from_predict(self, a, center, scale):
      """
      Extract points from the model prediction heatmaps
      """
      a_ch, a_h, a_w = a.shape

      print(f"get_pts_from_predict: Input shape = ({a_ch}, {a_h}, {a_w}), dtype = {a.dtype}")

      # Make sure we have the right number of channels (should be 68 for landmarks)
      if a_ch != 68:
         print(f"Warning: Expected 68 channels for landmarks, got {a_ch}. Using fallback landmarks.")
         base_landmarks = LandmarksProcessor.landmarks_2D
         # Create fallback landmarks based on center
         fallback_pts = base_landmarks.astype(np.float32) * 100 + np.array([center[0], center[1]], dtype=np.float32)
         return fallback_pts

      b = a.reshape((a_ch, a_h*a_w))
      c = b.argmax(1).reshape((a_ch, 1)).repeat(2, axis=1).astype(np.float32)  # Force to float32
      c[:,0] %= a_w
      c[:,1] = np.apply_along_axis(lambda x: np.floor(x / a_w), 0, c[:,1])

      for i in range(a_ch):
         pX, pY = int(c[i,0]), int(c[i,1])
         # Use dynamic boundary check instead of hardcoding 63
         if pX > 0 and pX < a_w-1 and pY > 0 and pY < a_h-1:
            diff = np.array([a[i,pY,pX+1]-a[i,pY,pX-1], a[i,pY+1,pX]-a[i,pY-1,pX]], dtype=np.float32)  # Force to float32
            c[i] += np.sign(diff)*0.25

      c += 0.5

      # Apply a scaling factor to adjust landmark size if needed
      pts = np.array([self.transform(c[i], center, scale, a_w) for i in range(a_ch)], dtype=np.float32)  # Ensure output is float32

      # Add a small offset to adjust for potential bias in the model's output
      # This is based on empirical observation that FAN model tends to produce slightly oversized landmarks
      offset_factor = 0.95  # Reduce by 5%

      # Calculate centroid of landmarks
      centroid = np.mean(pts, axis=0)
      # Scale distances from centroid by the offset factor
      pts = centroid + (pts - centroid) * offset_factor

      return pts

		自动登录	找回密码
密码			立即注册（仅限QQ邮箱）