SdaPS*: A Novel Source-Free Domain Adaption Method for Point Cloud Primitive Segmentation

As a crucial step in intelligent manufacturing, 3-D primitive segmentation enhances the efficiency of industrial automation by dividing surface point clouds into various geometric primitives. However, variations in object shapes across industrial datasets lead to domain shift issues and privacy concerns hinder dataset sharing, limiting the application of DNN-based primitive segmentation methods. To tackle these challenges, we propose SdaPS*, a novel source-free domain adaptation method for point cloud primitive segmentation to improve the generality in industrial scenarios, following self-training framework. SdaPS* utilize multiconfidence, including transformation consistency, task confidence, and geometric saliency, to provide more informative guidance. Initially, transformation consistency is leveraged to vote for robust initial pseudolabels. Then, a hierarchical label reassignment method is designed for label denoising, incorporating dynamic prototype distribution matching and label graph aggregation strategies for points with different task confidence. In addition, we propose a comprehensive self-training and inference strategy that leverages the coupled relationships between instances and semantics to achieve more accurate segmentation. Experimental evaluations verify the performance of our method and practical applications validate its effectiveness in industrial scenarios.