Yolov5s is not giving accurate detections after converting to onnx

[vmukund36]

@marcoslucianops I have converted my custom .pt model to .onnx using the export_yolov5.py file but when I run inference on deepstream(6.3) using Jetson AGX Orin, I get wrong detections and I dont get the correct ones also properly. If I infer it using Ultralytics, I am getting accurate results but not with deepstream. What could be the reason?

[nikola310]

Hi @vmukund36 how did you run inference with onnx? I tried using official yolov5 repo, but I can't get it to work since the output shapes are different from official version (official has one output, while the one I get with export_yolov5 has three outputs - classes, boxes & scores).

[marcoslucianops]

@vmukund36, can you send images of the output on DeepStream vs Torch?