Upload app.py

app.py

@@ -519,8 +519,13 @@ async def handle_upright_challenge(data: dict) -> dict:
         
         # Debug: Log image dimensions
         logger.info(f"🔍 UPRIGHT DEBUG: Original image dimensions: {original_w}x{original_h}")
-        
-        input_tensor = preprocess_image(image_bytes, model_data['input_size'])
+
+        # Determine input size for this inference (prefer model config, default to 640 for consistency with ONNX test)
+        input_size = model_data.get('input_size')
+        if not isinstance(input_size, int) or input_size <= 0:
+            input_size = 640
+
+        input_tensor = preprocess_image(image_bytes, input_size)
         outputs = model_data['session'].run(None, {model_data['input_name']: input_tensor})[0]
         
         predictions = np.squeeze(outputs).T
@@ -547,9 +552,9 @@ async def handle_upright_challenge(data: dict) -> dict:
         # Debug: Log model space coordinates
         logger.info(f"🔍 UPRIGHT DEBUG: Best detection (model space): x_center={box_model[0]:.2f}, y_center={box_model[1]:.2f}, width={box_model[2]:.2f}, height={box_model[3]:.2f}")
         
-        scale = min(model_data['input_size'] / original_w, model_data['input_size'] / original_h)
-        pad_x = (model_data['input_size'] - original_w * scale) / 2
-        pad_y = (model_data['input_size'] - original_h * scale) / 2
+        scale = min(input_size / original_w, input_size / original_h)
+        pad_x = (input_size - original_w * scale) / 2
+        pad_y = (input_size - original_h * scale) / 2
         
         # Debug: Log scaling parameters
         logger.info(f"🔍 UPRIGHT DEBUG: Scaling parameters: scale={scale:.4f}, pad_x={pad_x:.2f}, pad_y={pad_y:.2f}")