Segment all roads: Domain generalized freespace detection by robust surface normal information embedding and edge-aware learning

Abstract

Freespace detection (FD), also known as drivable area detection, plays a crucial role in autonomous driving. However, existing FD methods, which are based on supervised learning, struggle to perform well when testing on out-of-distribution data. Domain generalization is a prospective solution to address this issue, aiming to improve the generalization ability on out-of-distribution data (unseen domains). In this paper, we explore a new task, domain generalized freespace detection, and design a novel approach, which can generalize to both structured urban on-road scenes and unstructured off-road scenes. First, considering the robustness of surface normal (SN) information in diverse domains, we present a robust information embedding module to embed SN information into RGB features. Then, a cross-scale feature robustness enhancement module is introduced to aggregate RGB-SN features across different scales to obtain more domain-invariant features. Moreover, an RGB-SN edge-aware learning strategy is devised to further improve the generalization ability. Extensive experiments demonstrate that the proposed approach achieves superior performance compared to current approaches.