Few-Shot Point Cloud Semantic Segmentation via Support-Query Feature Interaction

Few-shot point cloud semantic segmentation is a challenging problem that aims to use only a limited number of labeled point clouds to recognize novel samples, significantly reducing the cost of manual annotation of point clouds. An effective solution is based on deep metric learning, by projecting points into an embedding space and constructing prototypes to calculate distance or similarity to query features. Since the intricate geometric attributes of point clouds, there is a substantial discrepancy between the distribution of query and support features, making the model incapable of distinguishing foreground and background points under similar geometric features. In this paper, we propose an efficient query-support feature alignment method for few-shot point cloud semantic segmentation. We incorporate the support (query) features containing interactive relationships to the query (support) features, effectively shortening the spatial distance between query and support samples. To further utilize the limited relationships and avoid introducing too much computation, we adopt the query point cloud prediction results as supervised information to reversely predict the category of the support prototypes. Extensive experiments on the S3DIS and ScanNet datasets show that our proposed methods significantly surpass the state-of-the-art methods under several N-way K-shot settings. The project is available at https://github.com/miniflash/SQFI.