tensorflow-bilateral-permutohedral

Port of permutohedral bilateral filtering to tensorflow as an op.

History of this permutohedral code:

• Adams, Andrew, Jongmin Baek, and Myers Abraham Davis. "Fast High‐Dimensional Filtering Using the Permutohedral Lattice." Computer Graphics Forum. Vol. 29. No. 2. Blackwell Publishing Ltd, 2010.
• Ported to mxnet https://github.com/piiswrong/permutohedral
• This port to Tensorflow

The mxnet port of the permutohedral lattice plays better with Tensorflow than Caffe ports (e.g. https://github.com/torrvision/crfasrnn) because the mxnet port was updated to CUDA streams.

Installation / Use

Currently only the GPU version is supported.

You will need to build Tensorflow from source; will get linker errors if you try to use the pip version of Tensorflow.

To build, just type "make". It uses a similar Makefile as Caffe.

When you build it, try running some tests, like "test/test_segment.py" or "test/test_slider_window.py".

Note:

This is ONLY supposed to do permutohedral lattice bilateral filtering, which is useful as a main part of the CRF-RNN algorithm.

inputs:

• image_feats: shape [batch, channels, rows, cols] i.e. NCHW shape. This is the image that is to be filtered.
• image_wrt: same shape as above, except possibly different number of channels. When filtering, distances between points are computed with respect to this.
• standard deviation of spatial part of filters

The images will be filtered with a (2+channels)-dimensional filter.