AttenMLP: An Attention-Enhanced Multilayer Perceptron for 3D Point Cloud Classification

Abstract

Three-dimensional (3D) point cloud classification is a fundamental task in computer vision, with critical applications in autonomous vehicles, robotics, and augmented reality. Despite significant progress, existing methods often struggle with the irregular spacing, varying densities, and diverse distributions inherent in real-world 3D data. This paper presents AttenMLP, a new Multilayer Perceptron architecture designed to robustly capture both local and global geometric features for enhanced 3D point cloud classification. AttenMLP employs a dual-pathway MLP configuration augmented with a sinusoidal positional encoding scheme and an efficient multi-head attention mechanism. The dual pathways process features in complementary sparse and dense representations, enabling the network to model diverse geometric patterns effectively. Positional encoding preserves crucial spatial context throughout the network, while multi-head attention facilitates multi-scale feature integration, improving the detection of both fine-grained and coarse geometric patterns. Evaluated on the ModelNet40 benchmark, AttenMLP achieves a Mean Class Accuracy (mAcc) of 92.4% and an Overall Accuracy (OA) of 93.85%, compared to several point cloud classification methods, while requiring 3.15MB of parameters and 1.85GFLOPs, demonstrating the viability of efficient attention augmented MLP architectures for practical 3D vision.