Han Xu ; Tang Yuan ; Wang Zhaoxuan ; Li Xianzhi

Abstract

Existing Transformer-based models for point cloud analysis suffer fromquadratic complexity, leading to compromised point cloud resolution andinformation loss. In contrast, the newly proposed Mamba model, based on statespace models (SSM), outperforms Transformer in multiple areas with only linearcomplexity. However, the straightforward adoption of Mamba does not achievesatisfactory performance on point cloud tasks. In this work, we presentMamba3D, a state space model tailored for point cloud learning to enhance localfeature extraction, achieving superior performance, high efficiency, andscalability potential. Specifically, we propose a simple yet effective LocalNorm Pooling (LNP) block to extract local geometric features. Additionally, toobtain better global features, we introduce a bidirectional SSM (bi-SSM) withboth a token forward SSM and a novel backward SSM that operates on the featurechannel. Extensive experimental results show that Mamba3D surpassesTransformer-based counterparts and concurrent works in multiple tasks, with orwithout pre-training. Notably, Mamba3D achieves multiple SoTA, including anoverall accuracy of 92.6% (train from scratch) on the ScanObjectNN and 95.1%(with single-modal pre-training) on the ModelNet40 classification task, withonly linear complexity. Our code and weights are available athttps://github.com/xhanxu/Mamba3D.