Point Cloud Segmentation with Guided Sampling and Continuous Interpolation

Gaoyang Zhang, Xinguo Liu

Abstract: Sampling and interpolation are pivotal in the design of 3D neural networks. Presently, farthest point sampling and \(k\)-NN interpolation are the predominant techniques. Nonetheless, the former can lead to information loss in feature-rich regions, while the latter might introduce noticeable discontinuities, compromising neural network performance. In this research, we address information loss with a novel method, DistrFPS, that considers the input information distribution during the farthest point sampling. Leveraging DistrFPS, we introduce a guided sampling module to retain crucial information for subsequent network layers. We also propose a continuous interpolation module grounded in barycentric interpolation to ensure spatial coherent feature propagation to higher resolution network layers. Our approach’s efficacy in preserving information is demonstrated empirically through signal reconstruction in both 2D and 3D realms. Comprehensive experiments on S3DIS, ScanNet, and ShapeNetPart affirm the advantages of our technique for point-based networks.