Multi-objective Evaluation of Deep Learning Based Semantic Segmentation for Autonomous Driving Systems

Recent applications of deep learning (DL) architectures for semantic segmentation had led to a significant development in autonomous driving systems (ADS). Most of the semantic segmentation applications for ADS consider a plethora of classes. Nevertheless, we believe that focusing only on the segmentation of drivable roads, sidewalks, traffic signs, and cars, can drive the improvement of navigation and control techniques in autonomous vehicles. In this study, some state-of-the-art topologies are analyzed to find a strategy that can achieve a uniform performance for the four classes. We propose a multiple objective evaluation method with the purpose of finding the non-dominated solutions in different DL architectures. Numerical results are shown using CityScapes, SYNTHIA, and CamVid datasets.