Deep Learning on Superpoint Generation with Iterative Clustering Network

In 3D point clouds, superpoint is a set of points that share common characteristics. Semantically pure superpoints can greatly reduce the number of points while ensuring that the points located in the same superpoint have common semantic information. In this paper, we propose an end-to-end method for generating semantically pure superpoints. Specifically, we first use a light PointNet-liked network to embed low-dimensional point clouds into feature space to obtain semantic information. Next, we use farthest point sampling (FPS) to sample K points as the initial cluster centers. For each center, we cluster the points by jointly considering spatial and feature space. After clustering, we update the feature of each cluster center by simply averaging the point feature in the same cluster. By iteratively clustering and updating the feature of clusters, we obtain coarse superpoints, which contain a few points incorrectly clustered. Finally, to eliminate incorrectly clustered points, we leverage the breadth-first-search (BFS) to find and fuse them to obtain fine superpoints, leading to improvement on semantically pure superpoints. Extensive experiments conducted on S3DIS and ScanNet demonstrate the effectiveness of the proposed method. Furthermore, we achieve the state-of-the-art on both two datasets.