search

Algolzw / image-restoration-sde

Image Restoration with Mean-Reverting Stochastic Differential Equations, ICML 2023. Winning solution of the NTIRE 2023 Image Shadow Removal Challenge.
https://algolzw.github.io/ir-sde/index.html
MIT License
559 stars 41 forks source link

Shadow removing model #37

Open heorhiikalaichev opened 1 year ago

heorhiikalaichev commented 1 year ago

Hi!

I can't find a pretrained model for the shadow removing task in the list you provided.

https://drive.google.com/drive/folders/1RxkJ9u-vR3UlJ5PYy1D7jienEZd5d6my

Algolzw commented 1 year ago

Hi here is the link for our Refusion model’s weights and results: https://drive.google.com/drive/folders/1Xy7z9CH0weck775ixePnjD6fZIjAbre7?usp=share_link

heorhiikalaichev commented 1 year ago

Thank you so much!

Is it possible to provide a tile processing script for deshadow task? I have out of memory when trying to process 3072*4608 image.

Algolzw commented 1 year ago

Hi, I think the current shadow removal code can't process such a large-size image. One possible way is to downsample the image with a smaller size (e.g., [w/2, h/2]) and then upsample it after deshadowing. It might lose some accuracy but is the simplest way to process large images.

And another choice is to re-train a latent-refusion model to process large-size images (please refer to the code of unet-latent and latent-dehazing).