Unpaired Image-to-Image Translation using Adversarial Consistency Loss

Yihao Zhao     Ruihai Wu     Hao Dong*    
(*: corresponding author)

Peking University     European Conference on Computer Vision (ECCV) 2020

[ArXiv Preprint] [Code] [BibTex]
Abstract

Unpaired image-to-image translation is a class of vision problems whose goal is to find the mapping between different image domains using unpaired training data. Cycle-consistency loss is a widely used constraint for such problems. However, due to the strict pixel-level constraint, it cannot perform shape changes, remove large objects, or ignore irrelevant texture. In this paper, we propose a novel adversarial-consistency loss for image-to-image translation. This loss does not require the translated image to be translated back to be a specific source image but can encourage the translated images to retain important features of the source images and overcome the drawbacks of cycle-consistency loss noted above. Our method achieves state-of-the-art results on three challenging tasks: glasses removal, male-to-female translation, and selfie-to-anime translation.

Unpaired Image-to-Image Translation

Figure 1. Example results of our ACL-GAN and baselines. Our method does not require cycle consistency, so it can bypass unnecessary features. Moreover, with the proposed adversarial-consistency loss, our method can explicitly encourage the generator to maintain the commonalities between the source and target domains.

Adversarial-Consistency Loss

Figure 2. The comparison of adversarial-consistency loss and cycle-consistency loss. The blue and green rectangles represent image domains S and T, respectively. Any point inside a rectangle represents a specific image in that domain.

Qualitative Results

Figure 3. Comparison against baselines on glasses removal.


Figure 4. Comparison against baselines on male-to-female translation.


Figure 5. Comparison against baselines on selfie-to-anime translation.

Quantitative Comparisons

Figure 6. Quantitative Comparisons to Baseline Methods. We show quantitative comparisons between our algorithm and the baseline methods.

Acknowledgements

This work was supported by the funding for building AI super-computer prototype from Peng Cheng Laboratory (8201701524), start-up research funds from Peking University (7100602564) and the Center on Frontiers of Computing Studies (7100602567). We would also like to thank Imperial Institute of Advanced Technology for GPU supports.