VectorMamba: Enhancing point cloud analysis through vector representations and state space modeling

Abstract

Point cloud data, despite its widespread adoption, poses significant challenges due to its sparsity and irregularity. Existing methods excel in capturing complex point cloud structures but struggle with local feature extraction and global modeling. To address these issues, we introduce VectorMamba, a novel 3D point cloud analysis network. VectorMamba employs a Vector-oriented Set Abstraction (VSA) method that integrates scalar, rotation, and scaling information into vector representations, enhancing local feature representation. Additionally, the Flash Residual MLP (FlaResMLP) module improves generalization and efficiency by leveraging anisotropic functions and explicit positional embeddings. To address global modeling challenges, we propose the PosMamba Block, a state-space-based module that incorporates positional encoding to preserve spatial information and mitigate the loss of geometric context in deeper layers. Experimental results on the ModelNet40 classification dataset, ShapeNetPart part segmentation dataset, and S3DIS semantic segmentation dataset demonstrate that VectorMamba outperforms baseline methods and achieves competitive performance compared to other approaches. The code and dataset are openly available at github.com/Shadow581/VectorMamba.