Li-Chongyi/Zero-DCE_extension

Images are saturated for adobe5k

Opened this issue · 2 comments

deepakkupanda commented

Hi,

I trained ZDCE++ in adobe5k for 100 epochs. I had kept 4000 images for training, 500 for validation and 500 for testing. I used the original folder for training. I found the results to be saturated.
One more question is on why we are not checking for validation loss. If we monitor validation loss, can this be improved.
Any help will be appreciated.

Thank you

a0014-WP_CRW_6320

a0001-jmac_DSC1459

a0002-dgw_005

deepakkupanda commented

Hi
Let me know if I am doing something wrong. Any help will be appreciated.
Thank you

MoeLightOvO commented

Maybe you should take some over-exposed images to trian your model. By the way, have you encountered noise problems? The model I trained has very obvious noise. I don't know why this happened.