Transformer-based Geometric Point Cloud Compression with Local Neighbor Aggregation

Recently, point cloud processing is becoming popular in AI-driven areas as 3D scanners are developing rapidly. However, this kind of data can have a massive file size, causing significant file storage and transmission difficulties. Compressing point clouds is challenging due to the disordered, sparse, and irregular point cloud structures. Therefore, there is a growing need to develop effective methods to compress point clouds while preserving their information. So far, many methods based on voxel and octree structures have been reported. However, these methods suffer from the information loss issue of local details at early stages, especially during the down-sampling step. In addition, while the global attention mechanism of Transformers has strength in capturing long-range dependency features, it has limitations in capturing local geometry position details. To address these issues, we propose a Transformer-based point cloud geometric compression method with a local neighbor aggregation module to preserve local spatial features during compression. Our method is based on the architecture of the autoencoder, and a Local Neighbor Aggregation module will address the local feature-capturing limitations of the global attention and local spatial data loss in Transformers. Compared with other methods, our method achieves an average of 30.49% and 23.67% bitrate savings in terms of PSNR DI and PSNR D2 respectively with a shorter decoding time.

I. Introduction

Learning-based point cloud compression has become a popular topic because of the increasing demands for the representation and transmission of high-detail point clouds in applications like autonomous driving, virtual reality, and 3D modeling. Formed by a collection of points, a point cloud is a widely-used 3D data representation. Each point has specific coordinate information representing the location of the point, associated with other attribute features such as normals, RGB color values, and density. Usually, point clouds can be modeled as point-based, voxel-based, and octree-based, which are prevalent in various point cloud compression methods including traditional methods and learning-based methods.