LTTPoint: A MLP-Based Point Cloud Classification Method with Local Topology Transformation Module

In recent years, advancements and enhancements in 3D scanning technology have significantly improved the accessibility of 3D data. Three-dimensional data serves a wide array of applications, including augmented reality, autonomous driving, and robotics. The precise and rapid classification of 3D point cloud data stands as a fundamental challenge within these domains. Nonetheless, the inherent attributes of point cloud data, such as its unordered structure and non-uniform density, present substantial hurdles in data processing. To tackle these obstacles, we have developed and implemented a point cloud classification model grounded in a residual Multi-Layer Perceptron (MLP) with an integrated local topological transformation module. This module assimilates geometric prior knowledge from the local neighborhood into the point cloud data and employs affine transformations to address the unordered and locally non-uniform density characteristics of point cloud data. The outcome is a succinct yet highly effective method for point cloud classification. Our approach’s competitiveness is demonstrated through experiments conducted on open-source datasets like ModelNet40 and ScanObjectNN, where it surpasses other recently proposed methods for point cloud classification.