non_max_suppression is very slow and doesn't appear to have a cuda or multi-threaded implementation

[andrei-pokrovsky]

It appears that tf.image.non_max_suppression currently takes about 200ms for about 8000 boxes, runs on a single CPU thread and doesn't have a GPU implementation.

Environment info

Operating System:
Ubuntu 16.04

Installed version of CUDA and cuDNN:
(please attach the output of ls -l /path/to/cuda/lib/libcud*):
8.0, 5.1.5

If installed from binary pip package, provide:

1. A link to the pip package you installed:
2. The output from python -c "import tensorflow; print(tensorflow.__version__)".
   0.12.0-rc1

[RobSalzwedel]

Is there any progress on this feature?

[aluo-x]

Also interested in this as well.

[harishannavajjala]

Are there any good examples online on how to use tf.image.non_max_suppression ?? And does using this in-built nms function over our own function improves performance of the model ?

[ppwwyyxx]

@harishannavajjala The documentation is pretty clear IMHO, which part do you have questions?
For speed the answer is always the same: it depends on the speed of "your own function" (which only you know) plus the overhead to transfer data between TF and your own function.

[machanic]

@ppwwyyxx also want GPU version of NMS? Does there any progress here?

[MaskVulcan]

@ppwwyyxx when I use nms. I set the profile and find mns is in CPU.Is there any way to use NMS on GPU?

[zhenglaizhang]

This feature will be fantastic

[jola6897]

In Work-efficient parallel non-maximum suppression for embedded GPU architectures the authors describe how to bring NMS to the GPU.

Also does anyone know what the difference is between the implementation of NonMaxSuppressionV2 and NonMaxSuppression in Tensorflow?

[keineahnung2345]

@jonla1 Someone has implemented a CUDA version based on the mentioned paper here:
https://github.com/jeetkanjani7/Parallel_NMS
And following is my PyCUDA version:
https://github.com/keineahnung2345/Parallel_NMS/tree/pycuda/PyCUDA

[zjjott]

+1

[joeyearsley]

https://github.com/zengarden/light_head_rcnn/tree/master/lib/lib_kernel/lib_nms_dev

Here is a GPU version which can be built for TF.

[SpaceInvader61]

There seems to be a related commit #28745

[github-actions]

This issue is stale because it has been open for 180 days with no activity. It will be closed if no further activity occurs. Thank you.