PointAS: an attention based sampling neural network for visual perception

Abstract

Harnessing the remarkable ability of the human brain to recognize and process complex data is a significant challenge for researchers, particularly in the domain of point cloud classification—a technology that aims to replicate the neural structure of the brain for spatial recognition. The initial 3D point cloud data often suffers from noise, sparsity, and disorder, making accurate classification a formidable task, especially when extracting local information features. Therefore, in this study, we propose a novel attention-based end-to-end point cloud downsampling classification method, termed as PointAS, which is an experimental algorithm designed to be adaptable to various downstream tasks. PointAS consists of two primary modules: the adaptive sampling module and the attention module. Specifically, the attention module aggregates global features with the input point cloud data, while the adaptive module extracts local features. In the point cloud classification task, our method surpasses existing downsampling methods by a significant margin, allowing for more precise extraction of edge data points to capture overall contour features accurately. The classification accuracy of PointAS consistently exceeds 80% across various sampling ratios, with a remarkable accuracy of 75.37% even at ultra-high sampling ratios. Moreover, our method exhibits robustness in experiments, maintaining classification accuracies of 72.50% or higher under different noise disturbances. Both qualitative and quantitative experiments affirm the efficacy of our approach in the sampling classification task, providing researchers with a more accurate method to identify and classify neurons, synapses, and other structures, thereby promoting a deeper understanding of the nervous system.