2020/12/09 by Zhiwei Jia, Jia, Zhiwei, Bodi Yuan +11 · 1 citation
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Multimodal Machine Learning Applications #cs.CV
paper · pdf · doi:10.48550/arxiv.2012.04932
Accepted to ICCV 2021
openalex publication_date 2020/12/09 · arxiv created 2021/10/06 · arxiv updated 2021/10/07 · openalex created_date 2021/10/11 · openalex updated_date 2026/07/28
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 particular, although producing visually reasonable outputs, the learned models usually flip the semantics of the inputs. To tackle this without using extra supervision, 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.