Learning Inter-superpoint Affinity for Weakly Supervised 3D Instance Segmentation

Due to the few annotated labels of 3D point clouds, how learn discriminative features clouds segment object instances is a challenging problem. In this paper, we propose simple yet effective instance segmentation framework that can achieve good performance by annotating only one for each instance. Specifically, tackle extremely segmentation, first oversegment cloud into superpoints in an unsupervised manner and extend point-level annotations superpoint level. Then, based on graph, inter-superpoint affinity mining module considers semantic spatial relations adaptively generate high-quality pseudo via semantic-aware random walk. Finally, volume-aware refinement applying volume constraints objects clustering graph. Extensive experiments ScanNet-v2 S3DIS datasets demonstrate our method achieves state-of-the-art weakly supervised task, even outperforms some fully methods. Source code available at https://github.com/fpthink/3D-WSIS .