PointDKT: Dual-Key Transformer for Point Cloud

Benefiting from the development of transformer-based networks in natural language processing and computer vision, researches on 3D point cloud understanding have made great progress in recent years. However, existing works often focus solely on aggregating local features or construct global dependencies through methods that significantly increase memory and computation complexity. Addressing these challenges, we present Point Dual-Key Transformer(PDKT), a novel and straightforward end-to-end network architecture. To expand the receptive fields and preserve the local geometric structural information, we introduce an efficient dual-key cross attention mechanism, which builds upon standard cross attention to facilitate global feature capture across the network without neglecting local detail. Furthermore, we adopt a learnable operator to automatically integrate local and global branches. Extensive experiment results demonstrate the effectiveness and superiority of our method on point cloud understanding tasks

I. Introduction

With the acquisition of 3D data becoming increasingly facile, its application has expanded across numerous fields, such as autonomous driving, medical treatment, and virtual reality. Consequently, the 3D point cloud understanding task have gained more attention.