Improving RGB-Thermal Semantic Scene Understanding With Synthetic Data Augmentation for Autonomous Driving

Abstract

Semantic scene understanding is an important capability for autonomous vehicles. Despite recent advances in RGB-Thermal (RGB-T) semantic segmentation, existing methods often rely on parameter-heavy models, which are particularly constrained by the lack of precisely-labeled training data. To alleviate this limitation, we propose a data-driven method, SyntheticSeg, to enhance RGB-T semantic segmentation. Specifically, we utilize generative models to generate synthetic RGB-T images from the semantic layouts in real datasets and construct a large-scale, high-fidelity synthetic dataset to provide the segmentation models with sufficient training data. We also introduce a novel metric that measures both the scarcity and segmentation difficulty of semantic layouts, guiding sampling from the synthetic dataset to alleviate class imbalance and improve the overall segmentation performance. Experimental results on a public dataset demonstrate our superior performance over the state of the arts.