debug x2

backend/engine.py

@@ -110,13 +110,14 @@ def run(model, video_path, line, config, on_frame, save_annotated=False, annotat
     results = model.track(
         source=video_path,
         tracker="bytetrack.yaml",
-        imgsz=config["imgsz"],
+        imgsz=736,    # MUST match OpenVINO export imgsz
         conf=config.get("conf", 0.12),
         iou=config.get("iou", 0.6),
         vid_stride=stride,
         stream=True,
         verbose=False,
-        persist=False  # MUST be False — True causes ByteTracker state leak across runs
+        persist=False, # MUST be False — True causes ByteTracker state leak across runs
+        batch=2       # MUST match OpenVINO export batch size
     )
 
     a = line[0]

backend/server.py

@@ -39,8 +39,8 @@ async def lifespan(app: FastAPI):
     # Warm up: run a dummy inference so OpenVINO compiles its graph now,
     # not on the first real user request
     try:
-        dummy = np.zeros((640, 640, 3), dtype=np.uint8)  # HWC — what YOLO predict() expects
-        model(dummy, verbose=False)
+        dummy_img = np.zeros((736, 736, 3), dtype=np.uint8)
+        model([dummy_img, dummy_img], verbose=False)  # list of 2 imgs triggers batch=2
         print("[BACKEND] Model warm-up complete.")
     except Exception as e:
         print(f"[BACKEND] Warm-up skipped: {e}")