Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics

TL;DR

This paper introduces a method for unpaired image translation that maintains semantic consistency by enforcing robustness to input variations, thereby reducing semantics flipping without extra supervision.

Contribution

It proposes a semantic robustness loss based on multi-scale feature perturbations to improve semantic preservation in unpaired image translation.

Findings

Reduces semantics flipping in translated images.

Outperforms existing methods quantitatively and qualitatively.

Maintains semantic invariance without extra supervision.

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 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.