What make non_max_suppression fast?

[yphola]

In some earlier version non_max_suppression was very slow when the model inferenced a batch with 32 images. In this version v9.1 it become very fast. What make non_max_suppression fast?
I convert python code x = x [XC [Xi]] to c++ code x = x.index(xc[xi]) in non_max_suppression, it is slow in the first loop and fast in the next. How to do this with libtorch?

[glenn-jocher]

Updates make it fast :)

In previous versions we attempted add-ons like merge-NMS as well with native pytorch code that slowed down NMS, now we are using torchvision NMS only, which is heavily optimized for cpu and cuda ops.

I'm not sure how to do the same with libtorch (no experience with it). Are you saying libtorch models do not have any NMS function available?

[yphola]

Thank you for your reply. I updated my C++ code according to your python code, and now it's much faster.