Enhancing Group Attention for Off-Road Semantic Segmentation via Transition-Aware Refinement

초록

We propose EGFormer, a semantic segmentation framework for robust terrain understanding in off-road autonomous driving. In unstructured environments, subtle transition zones between terrain types critically influence drivability, making reliable modeling of these regions essential. Building upon a group-attention-based segmentation backbone, EGFormer introduces a transition-aware refinement mechanism that explicitly enhances feature representations in terrain transition bands, enabling more reliable segmentation of gradually blending surfaces. Experiments on the RUGD and RELLIS-3D datasets show that while overall mean Intersection over Union (mIoU) improves modestly, boundary IoU (bIoU), which more directly reflects prediction quality near mixed terrain transitions, achieves more substantial gains. Under data-limited training conditions, EGFormer exhibits reduced performance degradation compared to baseline methods, indicating improved robustness. Qualitative results further demonstrate more stable and semantically consistent predictions in ambiguous regions, even under noisy or incomplete annotations. These results suggest that transition-focused feature refinement improves reliability in the ambiguous transition regions most critical to off-road terrain understanding.

키워드