search

jasonbunk / tensorflow-bilateral-permutohedral

Port of permutohedral bilateral filtering to tensorflow as an op
Apache License 2.0
22 stars 4 forks source link

Any progress on implementing CRF RNN in tensorflow? #1

Open gaopeng-eugene opened 7 years ago

gaopeng-eugene commented 7 years ago

The receptive field of bilateral filter is the whole image. Do you have any suggestion on implementing an window bilateral filter.

gaopeng-eugene commented 7 years ago

Where is the describe function in your test code?

gaopeng-eugene commented 7 years ago

can you provide the trained_segment_weights.pkl file?

jasonbunk commented 7 years ago

About the receptive field, in theory a Gaussian has an infinite receptive field, but in the algorithm it is approximated by a series of filters on a lattice that are only 3 lattice points wide (the filter coefficients are just 1,2,1 on each lattice axis). You can change the spatial window size with the argument "stdv_space". Details about the lattice are in the paper: http://graphics.stanford.edu/papers/permutohedral/permutohedral.pdf

About the describe function, I added it to test_utils, it just prints debug info about a variable.

jasonbunk commented 7 years ago

The trained_segment_weights.pkl file is generated by test/test_segment.py after letting it train and keyboard-interrupting it to stop the training. I pushed a commit with the describe function.