Unsupervised Domain Adaption of Multimodal Traffic Scenes Point Cloud Semantic Segmentation based on Self-Attention

Unsupervised domain-adaptive (UDA) semantic segmentation of 3D LiDAR point clouds in multimodal traffic scenes faces challenges in segmenting small objects like pedestrians and cyclists, with cross-modal learning often leading to misleading results. To address this problem, the paper proposes TransUDA, a multi-modal point cloud semantic segmentation network based on transformers and multi-scale information fusion. The model improves segmentation of small objects by leveraging an effective receptive field. Additionally, a cross-modal learning mechanism with entropy adaptive weight redistribution ensures proper guidance of inter-modal information. Results show that TransUDA achieves a 2.7% mIoU improvement in Nuscenes (USA-Singapore) and 1.38% mIoU in DayNight, with 3.5% IoU improvement for pedestrians and 7.89 % IoU improvement for cyclists.