Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics
Abstract
Many applications of unpaired image-to-image translation require the input contents to be preserved semantically during translations. Unaware of the inherently unmatched semantics distributions between source and target domains, existing distribution matching methods (i.e., GAN-based) can give undesired solutions. In specific, although producing visually reasonable outputs, the learned models usually flip the semantics of the inputs. To tackle this without using extra supervisions, we propose to enforce the translated outputs to be semantically invariant w.r.t. small perceptual variations of the inputs, a property we call ""semantic robustness"". By optimizing a robustness loss w.r.t. multi-scale feature space perturbations of the inputs, our method effectively reduces semantics flipping and produces translations that outperform existing methods both quantitatively and qualitatively.
Cite
Text
Jia et al. "Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics." International Conference on Computer Vision, 2021. doi:10.1109/ICCV48922.2021.01401Markdown
[Jia et al. "Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics." International Conference on Computer Vision, 2021.](https://mlanthology.org/iccv/2021/jia2021iccv-semantically/) doi:10.1109/ICCV48922.2021.01401BibTeX
@inproceedings{jia2021iccv-semantically,
title = {{Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics}},
author = {Jia, Zhiwei and Yuan, Bodi and Wang, Kangkang and Wu, Hong and Clifford, David and Yuan, Zhiqiang and Su, Hao},
booktitle = {International Conference on Computer Vision},
year = {2021},
pages = {14273-14283},
doi = {10.1109/ICCV48922.2021.01401},
url = {https://mlanthology.org/iccv/2021/jia2021iccv-semantically/}
}