Upload main.py

main.py

@@ -61,13 +61,13 @@ def extract_clip(video_path, start_frame):
     """
     Open a fresh cap, seek ONCE to start_frame, then read frames
     SEQUENTIALLY (no cap.set inside loop). This is reliable for all codecs.
-    Collects SEQUENCE_LENGTH face crops using bounding box caching for speed.
+    Collects SEQUENCE_LENGTH face crops using downscaled frame-by-frame tracking.
     """
     cap = cv2.VideoCapture(video_path)
     cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)  # Seek exactly ONCE
 
     faces    = []
-    base_box = None
+    last_box = None
     attempts = 0
 
     while len(faces) < SEQUENCE_LENGTH and attempts < 60:
@@ -79,25 +79,33 @@ def extract_clip(video_path, start_frame):
         frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
         pil_img   = Image.fromarray(frame_rgb)
 
-        if base_box is None:
-            # Downscale for faster MTCNN detection
-            detect_img = pil_img.copy()
-            if detect_img.width > 640:
-                ratio = 640.0 / detect_img.width
-                detect_img = detect_img.resize((640, int(detect_img.height * ratio)))
-                boxes, _ = mtcnn.detect(detect_img)
-                if boxes is not None and len(boxes) > 0:
-                    base_box = [b / ratio for b in boxes[0].tolist()]
-            else:
-                boxes, _ = mtcnn.detect(pil_img)
-                if boxes is not None and len(boxes) > 0:
-                    base_box = boxes[0].tolist()
+        # Downscale for faster MTCNN detection on every frame
+        detect_img = pil_img.copy()
+        current_box = None
         
-        if base_box is not None:
-            # We use a larger padding (0.5) because we are caching the box and the face might move slightly
-            crop = crop_face(pil_img, base_box, padding=0.5)
-            if crop is not None:
-                faces.append(crop)
+        if detect_img.width > 640:
+            ratio = 640.0 / detect_img.width
+            detect_img = detect_img.resize((640, int(detect_img.height * ratio)))
+            boxes, _ = mtcnn.detect(detect_img)
+            if boxes is not None and len(boxes) > 0:
+                current_box = [b / ratio for b in boxes[0].tolist()]
+        else:
+            boxes, _ = mtcnn.detect(pil_img)
+            if boxes is not None and len(boxes) > 0:
+                current_box = boxes[0].tolist()
+
+        if current_box is not None:
+            smoothed = smooth_box(current_box, last_box)
+            last_box = smoothed
+        elif last_box is not None:
+            smoothed = last_box   # Hold last known position
+        else:
+            continue              # No face yet — keep reading
+
+        # Use standard padding for centered faces
+        crop = crop_face(pil_img, smoothed, padding=0.35)
+        if crop is not None:
+            faces.append(crop)
 
     cap.release()