search

Fediory / HVI-CIDNet

"You Only Need One Color Space: An Efficient Network for Low-light Image Enhancement"
MIT License
190 stars 16 forks source link

Challenges in Enhancing Dark Regions within ROI Using HVI-CIDNet on Smartphone-Captured Images #23

Closed DevilANANDGupta closed 3 months ago

DevilANANDGupta commented 4 months ago

Hey @Fediory ,

Thank you for providing the HVI-CIDNet model. I have tested it on the LOL dataset for entire image enhancement, but I am facing challenges when trying to enhance only the ROI (Region of Interest). No model seems to perform well in this specific case, as the ROI pixels appear significantly different from the rest of the image.

Could you please suggest an approach for enhancing dark regions only within the ROI? Additionally, I am working with a dataset consisting solely of smartphone-captured images, without any paired images. What would be the best method for low-light and normal-light image conversion in this context? How can I fine-tune the model for my dataset effectively?

Your guidance is crucial. Thank you

Fediory commented 4 months ago

Thank you so much for your question, it made me think twice about our low-light enhancement methods. Next I will answer your questions mainly based on our HVI-CIDNet.

Q1: What would be the best method for low-light and normal-light image conversion in this context? A1: Currently publicly available unsupervised models are better suited to unpaired datasets (in your case). For models such as CIDNet that require paired datasets for training, there are two ways to improve the quality of recovery.

Q2: How can I fine-tune the model for my dataset effectively? A2: The first approach, modify the CIDNet loss and framework to an unsupervised method and train on your dataset. The second approach, if still using weights trained on LOL, involves finding the differences in the camera that captured the LOL dataset versus the camera of your dataset, which is one of the key factors affecting the final generalisation ability. The greater the camera difference, the greater the difference in generalisation ability on other unpaired datasets. If this difference in generalisation ability due to camera differences is to be boosted, this difference needs to be customised using the T(x) function and the γ parameter in the paper (which can be interpreted as a translation of your unpaired dataset photographs into the LOL dataset's distribution)