Wrong detection with custom YOLOv5 RKNN unlike the published YOLOv5s model that works fine

[Guemann-ui]

Hi all,

I manage to get my custom RKNNs running on RK3588, but when I added the post-processing step, I got overlapped BBoxes on the whole output image, unlike the YOLOv5s that works fine and provide the inference correctly on the image. I am wondering to know if anyone has the steps how the published YOLOV5s onnx converted from PyTorch to onnx as I used the necessary steps to convert the model to onnx and then to rknn and everything works fine on RK3588 except the postprocessing.

Thanks

[yanrihong]

i have the same matters have u solved it ?

[Guemann-ui]

@yanrihong yes, you should cut the last three Conv layers