Workaround problem of tensorrt hanging sometimes.

extras/mmdeploy/configs/detection_onnxruntime_static-320x320_fp16.py

@@ -5,5 +5,14 @@ onnx_config = dict(
 )
 
 codebase_config = dict(
-    post_processing=dict(score_threshold=0.3, iou_threshold=0.3),
+    # For later TensorRT inference, the number of output boxes needs to be as stable as possible, 
+    # because a drop in the box count leads to a re-optimization which takes a lot of time,
+    # therefore reduce the maximum number of output boxes to the smallest usable value and sort out
+    # low confidence boxes outside the model.
+    post_processing=dict(
+        score_threshold=0.0,
+        confidence_threshold=0.0,
+        iou_threshold=0.3,
+        max_output_boxes_per_class=10,
+    ),
 )