what's the original NMS method in yolov3？

[Gaondong]

I saw that you have made many efforts on NMS.
What is the original yolov3 paper nms speed？ Did it in this project? I want to compare them.
I am failed install torchvision, So I test fast-NMS on jetson tx2, It cost 10ms.
What's difference in torchvision method and original yolov3 nms method.

[glenn-jocher]

@Gaondong torchvision nms and yolov3 nms are the same.

[Gaondong]

@Gaondong torchvision nms and yolov3 nms are the same.
@glenn-jocher
It sames that the torchvision method is faster and more accuracy than Fast-nms.
But in Fast-nms paper, It sames original nms methods cost more time.

[glenn-jocher]

See #679

Fast NMS is slower than torchvision.

[Gaondong]

See #679

Fast NMS is slower than torchvision.
@glenn-jocher
Yes I saw your torchvision is best performance,
I have also used your previous version, nms is relatively slow. Is it just because the calculation method of torchvision is faster than the previous calculation of torch cuda?
Why is the torchvision version so much faster than the original yolov3 paper implementation？

[glenn-jocher]

Torchvision has cpu and CUDA optimized versions, I’m assuming they programmed them for optimal runtime on both sets of hardware.

[Gaondong]

Torchvision has cpu and CUDA optimized versions, I’m assuming they programmed them for optimal runtime on both sets of hardware.

Thanks for your reply.
In AleayAB's darknet project, there is a hyper parameter counter_per_class,
https://github.com/AlexeyAB/darknet/issues/4792
Could it in this project？

[glenn-jocher]

@Gaondong we don't have this, but if you'd like to submit a PR we can review it!

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.