attPCPR: attention-aware prototype cross-propagation and relocation for one-shot 3D semantic segmentation

Point cloud semantic segmentation significantly facilitates the understanding of driving environment and information of significance for autonomous vehicles. However, the huge amount of annotated data required for obtaining well-performed segmentation models via supervised training has impeded further development of point-wise semantic segmentation. In addressing this challenge, our study leverages few-shot learning to improve model performance while drastically reducing the need for labeled point clouds. We also identify substantial untapped potential in the learning capabilities of segmentation networks within a few-shot learning framework, particularly when encountering novel testing categories unseen during training. Integrating the concept of prototypical networks, our method facilitates the generation of multiple prototypes to effectively model point-wise data structures via a cross-propagation strategy designed to enhance the learning of vital information in support and query samples. Drawing inspiration from the success of attention mechanism across various domains, we introduce an attention relocation module to refine attention-aware feature learning in each episode. The implications of employing multiple attention heads are also explored to further augment our model's performance. To validate the effectiveness of our method, we conduct a series of experiments and ablation studies on the well-known benchmark datasets S3DIS and ScanNet. Compared to baselines and the state-of-the-arts, our approach achieves markedly improved one-shot segmentation performance. Subsequent experiments on more shots of data also showcase the effective generalization ability of our method.