First commit

app.py

@@ -65,7 +65,20 @@ def extract_faces(video_file):
         # Inference
         results = model(frame, verbose=False)[0]
         for box in results.boxes:
-            x1, y1, x2, y2 = map(int, box.xyxy[0])
+            # version 1
+            # x1, y1, x2, y2 = map(int, box.xyxy[0])
+            # version 2
+            h, w, _ = frame.shape
+            x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
+            pad_ratio = 0.3  # 30% padding
+
+            dx = (x2 - x1) * pad_ratio
+            dy = (y2 - y1) * pad_ratio
+
+            x1 = int(max(0, x1 - dx))
+            y1 = int(max(0, y1 - dy))
+            x2 = int(min(w, x2 + dx))
+            y2 = int(min(h, y2 + dy))
             face_crop = frame[y1:y2, x1:x2]
             if face_crop.size != 0:
                 resized = cv2.resize(face_crop, (224, 224))
@@ -77,9 +90,7 @@ def extract_faces(video_file):
     # Save as video
     tmpdir = tempfile.mkdtemp()
     output_path = os.path.join(tmpdir, "face_only_video.mp4")
-    #clip = ImageSequenceClip([cv2.cvtColor(f, cv2.COLOR_BGR2RGB) for f in frames], fps=fps)
-    #clip.write_videofile(output_path, codec="libx264", audio=False, verbose=False, logger=None)
-    clip = ImageSequenceClip(frames, fps=25)
+    clip = ImageSequenceClip([cv2.cvtColor(f, cv2.COLOR_BGR2RGB) for f in frames], fps=25)
     clip.write_videofile(output_path, codec="libx264", audio=False)
 
     return output_path