Few-Shot 3D Point Cloud Semantic Segmentation

Many existing approaches for point cloud semantic segmentation are strongly supervised. These strongly supervised approaches heavily rely on a large amount of labeled training data that is difficult to obtain and suffer from poor generalization to new classes. To mitigate these limitations, we propose a novel attention-aware multi-prototype transductive few-shot point cloud semantic segmentation method to segment new classes given a few labeled examples. Specifically, each class is represented by multiple prototypes to model the complex data distribution of 3D point clouds. Subsequently, we employ a transductive label propagation method to exploit the affinities between labeled multi-prototypes and unlabeled query points, and among the unlabeled query points. Furthermore, we design an attention-aware multi-level feature learning network to learn the discriminative features that capture the semantic correlations and geometric dependencies between points. Our proposed method shows a significant improvement compared to the baselines for few-shot point cloud segmentation on unseen classes in two benchmark datasets.