Home

Awesome

Official repository for the paper

Day-to-Night Image Synthesis for Training Nighttime Neural ISPs @CVPR'22 Oral

Abhijith Punnappurath, Abdullah Abuolaim, Abdelrahman Abdelhamed, Alex Levinshtein, and Michael S. Brown

Samsung Artificial Intelligence Center, Toronto, Canada

[Paper] [arXiv] [Supplemental] [Samsung Reserach blog post]

alt text

Citation

If you use this code or the associated data, please cite the paper:

@InProceedings{Punnappurath_2022_CVPR,
author = {Punnappurath, Abhijith and Abuolaim, Abdullah and Abdelhamed, Abdelrahman and Levinshtein, Alex and Brown, Michael S.},
title = {Day-to-night Image Synthesis for Training Nighttime Neural ISPs},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2022},
}

Note

Get started

Convert day images to synthetic nighttime images

Train Night Mode ISP on synthetic nighttime images

Test the synthetic nighttime data model

Test using our pretrained synthetic nighttime data model

ModelPSNR (dB)SSIM (dB)
Clean raw ISO 5045.970.9924
Noisy raw ISO 160037.000.9288
Noisy raw ISO 320036.140.9182

Train and test Night Mode ISP on real data

Test using our pretrained real data model

Results averaged over 3 folds

ModelPSNR (dB)SSIM (dB)
Clean raw ISO 5046.660.9923
Noisy raw ISO 160039.250.9511
Noisy raw ISO 320038.140.9406

Baseline models day, day_dimmed, and global_relight

Generate data

Train and test

Mixing synthetic and real data

(Optional) Full burst dataset:

Contact

Abhijith Punnappurath - (abhijith.p@samsung.com; jithuthatswho@gmail.com)