Effective Use of Synthetic Data for Urban Scene Semantic Segmentation

TL;DR

This paper proposes a novel method for training semantic segmentation models using only synthetic data, avoiding the need for real images during training by leveraging different handling of foreground and background classes.

Contribution

It introduces a detection-based approach for foreground classes to improve domain transfer, eliminating the need for real images during training.

Findings

Effective on Cityscapes and CamVid datasets

Outperforms traditional synthetic training methods

No real images needed during training

Abstract

Training a deep network to perform semantic segmentation requires large amounts of labeled data. To alleviate the manual effort of annotating real images, researchers have investigated the use of synthetic data, which can be labeled automatically. Unfortunately, a network trained on synthetic data performs relatively poorly on real images. While this can be addressed by domain adaptation, existing methods all require having access to real images during training. In this paper, we introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time. Our approach builds on the observation that foreground and background classes are not affected in the same manner by the domain shift, and thus should be treated differently. In particular, the former should be handled in a detection-based manner to better account for the fact that, while their texture in synthetic images is not photo-realistic, their shape looks natural. Our experiments evidence the effectiveness of our approach on Cityscapes and CamVid with models trained on synthetic data only.