MGSMamba: hierarchical grouped learning with shifted bidirectional state-space models for point cloud classification and segmentation
Point cloud classification is a fundamental task in 3D vision and robotics, yet effectively capturing local geometric structures while modeling global contextual dependencies remains challenging. Existing Mamba-based frameworks excel in sequential modeling but often underutilize spatial and structural information intrinsic to unordered3D point clouds. To address these limitations, we propose MGSMamba, a multi-stagepoint cloud classification framework integrating grouping structures with bidirectionalstate space modeling. Our framework implements a hierarchical three-level feature extraction process: intra-group structural modeling via a group-wise embedding module(GEM) incorporating local normals and point density, inter-group semantic interactionvia a cross-group feature learning module (CGLM) with grouped affine normalization(GANorm) and rotational feature-enhanced pooling (RFEP), and global contextual understanding through a shifted bidirectional state space module (Shifted Bi-SSM). Experiments on ModelNet40, ScanObjectNN, and ShapeNet-Part datasets demonstrate thatMGSMamba outperforms state-of-the-art methods in accuracy, robustness, and generalization, with a 2–4% improvement in classification accuracy across benchmarks whilepreserving invariance to point order. Our approach offers a scalable, interpretable, andreproducible framework for point cloud analysis with potential applications in robotics,autonomous navigation, and 3D scene understanding. The source code is available atGitHub: https://github.com/2359063436/MGSMamba, and detailed documentation isprovided in the repository for full reproducibility.
Authors
- Leping He (ORCID: https://orcid.org/0000-0002-2566-487X)
- Qijie Cai (ORCID: https://orcid.org/0000-0002-2269-1477)
- Xingpeng Zhang (ORCID: https://orcid.org/0000-0001-6547-6374)
- Qijun Hu (ORCID: https://orcid.org/0000-0002-9747-4311)
- Yu Bai (ORCID: https://orcid.org/0000-0001-5075-1505)
- Yang Yu (ORCID: https://orcid.org/0000-0002-0784-9715)
- Zengyi Zhang
- Qiang Zeng
Institutions
- Southwest Petroleum University (CN)
- Southwest Jiaotong University (CN)
Publication Details
- Journal
- The Visual Computer
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1007/s00371-026-04835-x
- Primary Topic
- 3D Shape Modeling and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00