A Survey on Transformers for Point Cloud Processing: An Updated Overview

Abstract

In recent years, the popularity of depth sensors and three-dimensional (3D) scanners has led to the rapid development of 3D point clouds. A transformer is a type of deep neural network mainly based on a self-attention mechanism, which was first applied in the field of natural language processing. Due to its strong capabilities for capturing contextual information, researchers are focusing on applying transformers to point cloud processing tasks. In various point cloud benchmarks, transformer-based models have exhibited similar or better performances than other types of networks, such as PointNet++ and PointCNN. Given their excellent performances, transformers have attracted growing attention in area of point cloud processing. In this paper, we review these point cloud transformer models by categorizing them in two main tasks: high-/mid-level vision and low-level vison, and we analyze their advantages and disadvantages. Furthermore, we also review the formulation of transformers and their base component: the self-attention mechanism. In addition, we compare common datasets for point cloud processing to help researchers select the most suitable datasets for their tasks. Finally, we discuss the challenges and briefly examine several future research directions for point cloud transformers.