Thanks for your implementation, I read your code and find some your implementation bugs about our cvpr work. The network's output should be an illumination map, you should use input image pixel divide the illumination map to get the predicted result. then calculate the L2 loss.
Local smoothness loss is not tv loss. You can find the implementation in our repo, metric.py. there are two parameters in the local smoothness loss function, the input is the original image, and output is the illumination map.
We also give a color loss implementation in our repo issue. You can have a try.
JIA-Lab-research/DeepUPE#45
Thanks for your implementation, I read your code and find some your implementation bugs about our cvpr work. The network's output should be an illumination map, you should use input image pixel divide the illumination map to get the predicted result. then calculate the L2 loss.
Local smoothness loss is not tv loss. You can find the implementation in our repo, metric.py. there are two parameters in the local smoothness loss function, the input is the original image, and output is the illumination map.
We also give a color loss implementation in our repo issue. You can have a try.
JIA-Lab-research/DeepUPE#45