Content Disentanglement for Semantically Consistent Synthetic-to-Real Domain Adaptation

  • 2021-08-03 15:55:28
  • Mert Keser, Artem Savkin, Federico Tombari
  • 0

Abstract

Synthetic data generation is an appealing approach to generate novel trafficscenarios in autonomous driving. However, deep learning perception algorithmstrained solely on synthetic data encounter serious performance drops when theyare tested on real data. Such performance drops are commonly attributed to thedomain gap between real and synthetic data. Domain adaptation methods that havebeen applied to mitigate the aforementioned domain gap achieve visuallyappealing results, but usually introduce semantic inconsistencies into thetranslated samples. In this work, we propose a novel, unsupervised, end-to-enddomain adaptation network architecture that enables semantically consistent\textit{sim2real} image transfer. Our method performs content disentanglementby employing shared content encoder and fixed style code.

 

Quick Read (beta)

loading the full paper ...