NMS in GPU

[MaskVulcan]

I find NMS operations is in CPU.Is there any way to switch to GPU?

[RobertCsordas]

As far as I know, TF has only a CPU based NMS implementation. I don't think so that it can be parallelized so well that it can get any significant performance gains on GPU.

[machanic]

@xdever I think if you calculate all pairs of boxes IOU first, then just for-loop once will ultimately boost speed, there have some trick in it, just see the source code in https://github.com/rbgirshick/py-faster-rcnn/tree/master/lib/nms
I think cuda version of NMS is faster than CPU version if we compile against tensorflow. The above code still have one issue if we use in tensorflow: it moves the data from cpu to gpu memory. We just need to define a new op and copy the source code of rbgirshick in it.