Repository navigation
Color shift #14
Description
Activity
Dear Mingzhen,
Thanks for your valuable observation.
We also encounter similar failure cases during test.
I think there are two possible reasons:
- The overall tonemapping style from the SICE dataset is not suitable for these cases.
- The decomposition strategy applied to the under exposure leads to the ignorance of the illuminance component.
So, maybe you can try following experiment to improve the performance over these failure cases:
- Add some data that contain similar scenes of these failure cases.
- Add loss function to force model to learn do not modify color in specific regions.
- Do not perform decomposition on under exposure (directly input the RGB image).
These are my initial thoughts. We can further discuss how to solve this problem elegantly.
Thank you very much! I'll try them.
Happy National Day!
Dear Authors，
Thank you for releasing this amazing project.
I found some issues for extreme EV. For example, color shift. I found that the image's saturation has decreased after recovery (compared to the ground truth), i.e., dark black turned light black .
Have you ever encountered this phenomenon?
Would you please share any ideas on how to solve it?
Sincerely
Mingzhen