Understanding Cross-Domain Robustness in LiDAR Semantic Segmentation

Abstract

Real-World deployment of perception models requires generalization beyond the environments encountered during training. However, collecting and annotating data that cover all possible conditions is infeasible. Consequently, models often suffer from performance degradation when applied to new domains, due to factors such as differences in beam configurations, sensor noise, and environmental conditions. Addressing these cross-dataset domain shifts is therefore essential for ensuring robustness and generalization. In this work, we evaluate perception model across domains and study strategies that help alleviate performance degradation.