Learning-to-See-in-the-Dark
Learning-to-See-in-the-Dark copied to clipboard
Doubt regarding the amplification ratio
The code is applying an amplification ratio as min(300, gt_exp / input_exp) and then doing a clip to reduce all values > 1 to 1.0
In some cases, this would cause the entire input packed bayer tensor to be all 1's.
Here's histogram of one such input tensor patch
Just wanted to clarify if this is okay and what is the intuition of processing an all 1's tensor through a CNN?
In this case, this means the GT image contains all ones, which is not in the dataset.
After the ratio is applied, the input should match the brightness of GT, while containing noise and quantization errors. It will not give you all ones if the GT is well-exposed.