import torch import numpy as np import cv2 import mediapipe as mp from torchvision import transforms # --- Torch image preprocessing (same as training) --- preprocess = transforms.Compose([ transforms.Grayscale(num_output_channels=3), transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize([0.5]*3, [0.5]*3) ]) # --- MediaPipe FaceMesh setup --- mp_face = mp.solutions.face_mesh face_mesh = mp_face.FaceMesh(static_image_mode=True, max_num_faces=1) def get_landmark_attention(image, size=224): """Compute attention map from MediaPipe face mesh landmarks and resize to match input size.""" h, w, _ = image.shape attn = np.zeros((h, w), dtype=np.float32) results = face_mesh.process(cv2.cvtColor(image, cv2.COLOR_RGB2BGR)) if results.multi_face_landmarks: for lm in results.multi_face_landmarks[0].landmark: x, y = int(lm.x * w), int(lm.y * h) if 0 <= x < w and 0 <= y < h: cv2.circle(attn, (x, y), 2, 1.0, -1) attn = cv2.GaussianBlur(attn, (31, 31), 0) attn = attn / np.max(attn) if np.max(attn) > 0 else attn attn_resized = cv2.resize(attn, (size, size)) return torch.tensor(attn_resized).unsqueeze(0).float() # shape (1,224,224)