SPS-LCNN: A Significant Point Sampling-based Lightweight Convolutional Neural Network for point cloud processing

Abstract

Point cloud data have very promising applications, but the irregularity and disorder make it a challenging problem how to use them. In recent years, an increasing number of new and excellent research solutions have been proposed, which focus on exploring local feature extractors. Over-engineered feature extractors lead to saturating the performance of current methods and often introduce unfavorable latency and additional overhead. This defeats the original purpose of using point cloud data, which is simplicity and efficiency. In this paper, we construct a learnable pipeline by designing two core modules with a small number of parameters – significant point sampling (SPS) and multiscale significant feature extraction (MS-SFE) – to balance accuracy and overhead. Our pipeline demonstrates comparable performance to state-of-the-art methods while requiring fewer parameters, making it well-suited for real-time applications.