SGSR: structure-guided semantic recovery for event-based spiking semantic segmentation

Event-based spiking semantic segmentation is well suited to sparse and asynchronous visual sensing, but dense prediction remains challenging because event responses are spatially non-uniform. Object contours and motion boundaries usually generate strong responses, whereas object interiors and regions with weak brightness changes provide limited evidence. This imbalance often results in incomplete semantic regions and inaccurate boundaries. To address this problem, we propose SGSR, a structure-guided spiking framework for event-based semantic segmentation. SGSR employs a multi-scale spiking separable convolution block to capture local structures at different spatial scales with low computational cost. A spiking parallel aggregation pyramid pooling module aggregates high-level context to improve the representation of weakly observed regions. In addition, a lightweight boundary-prior branch extracts boundary cues from intermediate features. The predicted priors are injected into successive decoder stages to guide spatial recovery, rather than being used only for auxiliary supervision. Under the event-only SNN setting, SGSR achieves mIoU scores of $$52.93%$$ on DDD17-Seg and $$52.45%$$ on DSEC-Semantic. The results show that SGSR provides competitive segmentation accuracy while maintaining a compact computational profile.