SCSA-Unet: cross-spatial positional feature aggregation for event-based semantic segmentation

Abstract

Event cameras offer considerable potential for autonomous driving perception owing to their high-dynamic range and microsecond-level temporal resolution. However, event representations tend to produce prominent responses around object boundaries and abrupt intensity changes while providing limited semantic cues within object interiors. This sparsity makes it difficult to recover complete and regionally consistent semantic predictions. To address this problem, we propose SCSA-Unet, an event-only semantic segmentation network designed for cross-spatial response organization and structural aggregation. Specifically, a cross-spatial positional interaction block is constructed to model associations among spatially separated valid activations, thereby improving the consistency of regional feature representations. A spatial-aware feature aggregation module is then introduced to aggregate multiscale contextual information at the deepest stage and strengthen high-level structural representation. Finally, a gated-guided fusion mechanism and mixed-scale boundary supervision are designed to refine decoder feature fusion and improve boundary localization. Experimental results on the DDD17-Seg and DSEC-Semantic benchmarks show that SCSA-Unet achieves mean intersection over union scores of 57.63% and 55.10%, respectively, supporting the effectiveness of the proposed event-only framework for semantic modeling under sparse event representations.