Implementation of Bilateral Guided Upsampling

[zakajd]

Hi, @mahmoudnafifi

While code for this project is not released I'd like to clarify which implementation of bilateral guided upsample did you use?
I found 3 versions:

1. Code from Google google/bgu.
2. Code from Google's HRNet paper google/hdrnet
3. Version from DeepUPE repo Jia-Research-Lab/DeepUPE.

All these options have their own limitations, 1 and 3 is pretty old and can't be compiled with new CUDA10+ and TF2+, while 2 is GPU only.

So my questions is: have you written your own custom implementation or adapted one of those?

[mahmoudnafifi]

Hi @zakajd,

The hdrnet and deepupe do the upsampling in their end-to-end training. This upsampling is not a part of our solution, I used it as a post-processing step in the testing phase only. That is why I used the non-learning one google/bgu. As this is a post-processing step and applied only in the testing phase, the CPU implementation is fine. Reimpelemnting it is on my to-do list but can't promise it.

[zakajd]

Thanks for the fast answer!
Did you use CUDA 10+ to compile the code from google/bgu?

[mahmoudnafifi]

No, I used the CPU Matlab implementation.