Active learning-driven semantic segmentation for railway point clouds with limited labels

Abstract

Accurate semantic segmentation of railway point clouds is crucial for railway infrastructure modelling. However, existing fully-supervised methods are heavily dependent on labeled datasets, while label-efficient methods typically struggle to generate representative annotations. To address these challenges, a weakly supervised point cloud semantic segmentation method based on active learning is proposed, significantly reducing labeling requirements while maintaining high segmentation accuracy. During the labeled pool updating phase, a strategy combining high-loss regions and high-uncertainty points is employed to actively select representative samples for annotation. To enhance the model's capacity for capturing complex railway structures, geometric features are embedded into the network encoder. Additionally, a class prototype dictionary is constructed, and dynamically weighted pseudo-labels are generated to maximize the utilization of limited supervisory information during training. Experimental results on three diverse railway datasets demonstrate that the method achieves superior segmentation accuracy with fewer labels compared to both popular weakly and fully supervised methods.