Learning-to-See-in-the-Dark
Learning-to-See-in-the-Dark copied to clipboard
cchen156
Can i prepare my own jpeg dataset and trait it ? instead of raw ?
Can i prepare my own JPEG or PNG dataset and trait it ? instead of raw ? Can i do this ? will it work ?
It is mentioned in the paper that they tried training it with jpeg images but the results were better with raw images as input indeed. However, you can easily try it yourself, no need to create a dataset, just postprocess short_exposed/input image as well like ground truth image. And the input channel will be reduced from 4 to 3.
It is mentioned in the paper that they tried training it with jpeg images but the results were better with raw images as input indeed. However, you can easily try it yourself, no need to create a dataset, just postprocess short_exposed/input image as well like ground truth image. And the input channel will be reduced from 4 to 3.
@ketankr9 could you clarify your suggestion? If I understand you correctly (and I am sorry if I misunderstood) what you are proposing here doesn't remove the need to feed RAW images through the network.
I think you are saying "run the raw images", but at the end convert the resulting RAW images to JPEG and get the results. If yes, then this would not be the same and is not what @Jayanth-L wants, which is to run his own JPEG images. If I am not mistaken if you want to use your own JPEG images, you have to come up with your own JPEG data set and then modify the code to have it handle JPEG. images. The existing code cannot handle JPEG images directly
To detect sensor noise you have to train the model on pixel level, you cant do that with lossy jpeg compression. @Jayanth-L
