Improving Binary Semantic Scene Segmentation for Robotics Applications

Abstract

Robotics applications are accompanied by particular computational restrictions, i.e., operation at sufficient speed, on embedded low power GPUs, and also for high-resolution input. Semantic scene segmentation performs an important role in a broad spectrum of robotics applications, e.g., autonomous driving. In this paper, we focus on binary segmentation problems, considering the specific requirements of the robotics applications. To this aim, we utilize the BiseNet model, which achieves significant performance considering the speed-segmentation accuracy trade-off. The target of this work is two-fold. Firstly, we propose a lightweight version of BiseNet model, providing significant speed improvements. Secondly, we explore different losses for enhancing the segmentation accuracy of the proposed lightweight version of BiseNet on binary segmentation problems. The experiments conducted on various high and low power GPUs, utilizing two binary segmentation datasets validated the effectiveness of the proposed method.