Few-shot 3D Point Cloud Semantic Segmentation with Prototype Alignment

Semantic Segmentation for 3D point clouds has made great progress in recent years. Most existing approaches for 3D point cloud segmentation are fully supervised, and they require a large number of well-annotated data for training. The training data is cost and quite difficult to obtain. Moreover, these fully supervised approaches cannot segment new classes well that are unseen in the training process. Thus, Few-shot segmentation has been developed to mitigate these limitations by learning to perform segment from a few labeled examples. In this paper, we propose a method to more adequately utilize information of query set and support set to promote performance of semantic segmentation for 3D point clouds. Specifically, we first extract support and query features and generate multiple prototypes to map the distribution of point clouds. Then we apply a transductive label propagation method to exploit the relations between labeled multi-prototypes and unlabeled points, and between pairs of unlabeled points. Finally, we utilize query points and predicted query masks to perform segmentation for support points. Our proposed method shows improvements for specific classes on S3DIS dataset compared to baselines in 2/3-way 1-shot point cloud semantic segmentation.