Debiasing Image-to-Image Translation Models
Md. Mehrab Tanjim (University of California, San Diego),* Krishna Kumar Singh (Adobe Research), Kushal Kafle (Adobe Research), Ritwik Sinha (Adobe Research), Garrison Cottrell (University of California, San Diego)
The 33rd British Machine Vision Conference
Abstract
Deep generative models have shown a lot of promise in various image-to-image translation tasks such as image enhancement and generating images from sketches. However, when all the classes are not equally represented in the training data, these algorithms can fail for underrepresented classes. For example, our experiments with CelebA-HQ face dataset reveal that this bias is prevalent for infrequent attributes, e.g., eyeglasses and baldness. Even when the input image clearly has eyeglasses, the image translation model is unable to create a face with them. To remedy this problem, we propose a data and model agnostic, general framework based on contrastive learning, re-sampling, and minority category supervision to debias existing image translation networks for various image-to-image translation tasks such as super-resolution and sketch-to-image. Our experimental results from the real and synthetic datasets show that our framework outperforms the baselines both quantitatively and qualitatively.Video
Citation
@inproceedings{Tanjim_2022_BMVC,
author = {Md. Mehrab Tanjim and Krishna Kumar Singh and Kushal Kafle and Ritwik Sinha and Garrison Cottrell},
title = {Debiasing Image-to-Image Translation Models},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0182.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by