issue related to detect multiple objects using the pose model

[xcn700418]

I successfully convert the model to onnx format and build the tensorRT engine. However, I am using the infer-pose.py to detect the object bounding boxes, each with associated keypoints, there are suppose to be 3 bounding boxes, but I only got 1 after nms. I check the coordinate of the detections that has higher scores than the conf_thres, they are quite similar, which means the model did not successfully detect the other object, but when i use the .pt model for inference, it did.

May I ask how can I modify the code to achieve multiple object detection? Is it related to the model conversion to onnx, or the tensorRT engine building stage, or related to the pose_postprocess function?