Miguel Monteiro

Hello, I have also developed a N-spatial-dimensional multi-channel version. [Check it out if you like](https://github.com/MiguelMonteiro/CRFasRNNLayer) Best Regards, Miguel

I found what I think is a bug in the implementation. The original permutohedral c++ [files](https://graphics.stanford.edu/papers/permutohedral/) suppose that you give your compute function value_size = image_channels + 1. In your...

I still would like to verify if @sadeepj can reproduce this bug or if I am just using it wrong. Nonetheless, I have a CPU version that works for bilateral...

I know, I have tried using your implementation with (channels, height, width) and instead of a blur I get a repeating structure overlaying the image (and possibly a blur underneath,...

For completion's sake, in your implementation when you switch (channel, height, width) for (height, width, channel) you get something like this: 

So in the first link `vd` is referred to as `value_size`. Regarding CRF Dense Inference your are not entirely correct. What you are referring to is `d` which is indeed...

I have invited you as a collaborator in my repo so you can see the code and bug for yourself. Ignore the master branch, the `sse` branch is your version...

I have finished a CPU and GPU version of the Lattice but I still have to get it into a Tensorflow kernel, I should finish it sometime around next week...

I haven't checked the dense3dCRF code but if the code from here came from there it might have the same bug. I decided to implement almost from scratch to avoid...

That's a good intuition of what it does, the permutohedral lattice is just the thing that does the bilateral filter provided you pass it the correct kernel. As for parameter...