ConUDA: Confidence-Guided Pseudo-Label Sampling for Unsupervised Domain Adaptation in 3D LiDAR Semantic Segmentation

Dense annotation of real 3D LiDAR point clouds for mobile robot applications remains challenging. Unsupervised Domain Adaptation (UDA) enables the segmentation of unlabeled real-world point clouds by leveraging labeled synthetic data. However, existing self-training-based UDA methods rely on fixed thresholds for pseudo-label selection, limiting adaptation performance. In this work, we address this limitation. We propose a novel UDA framework for 3D LiDAR semantic segmentation, centered on a confidence-guided pseudo-label sampling strategy (ConSamp). Specifically, ConSamp adopts a probabilistic sampling strategy in which pseudo-labels with higher confidence are more likely to be retained. Meanwhile, the sampling function itself evolves adaptively throughout training to respond to changes in confidence distribution. Experiments show that our model achieves strong performance on synthetic-to-real 3D LiDAR semantic segmentation tasks. In particular, results better than state-of-the-art methods have been achieved on two public 3D point cloud datasets: SemanticKITTI [1] and SemanticPOSS [2].