search

caiyuanhao1998 / Retinexformer

"Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement" (ICCV 2023) & (NTIRE 2024 Challenge)
https://arxiv.org/abs/2303.06705
MIT License
828 stars 64 forks source link

input_image after processing in Illumination_Estimator #56

Closed lucasta1 closed 5 months ago

lucasta1 commented 5 months ago

Hi,

In the paper, Eq. (5) is described as (Ilu,Flu)=E(I,Lp), Ien =R(Ilu,Flu). However, in RetinexFormer/basicsr/models/archs/RetinexFormer_arch.py, the output from E (illumination estimator) and the input of Denoiser formulated as a different way, as input_img = img * illu_map + img Link: https://github.com/caiyuanhao1998/Retinexformer/blob/448999c1462a8e3712dc6b08a9f3ec5a9ff32a12/basicsr/models/archs/RetinexFormer_arch.py#L337 Why did you describe as input_img = img * illu_map + img?

I'd be happy if you give me an explanation about it.

caiyuanhao1998 commented 5 months ago

the illumination map adjusts the illumination of the low-light image while the skip connection enables residual learning which helps ease the training process.

If you find our repo useful, please help us star it. Thanks.