GeoMamba: Dynamic Step-Size Adaptive Modulation for Mamba-Based Point Cloud Classification

Most current Mamba-based point cloud classification models use a constant step size, resulting in a fundamental limitation: the model fails to perceive the underlying structure of 3D geometric objects, thereby limiting classification accuracy. This limitation also renders the model susceptible to background noise in complex real-world environments. To address this issue, we propose GeoMamba. It introduces a lightweight, geometry-semantics dual-driven step-size modulation mechanism. By leveraging a multidimensional geometric encoder to extract relative positions, distance variations, and local neighborhood features, combined with semantic gating for context-aware modulation, our approach dynamically constrains the step-size adjustment range to [0.2, 5.0] times. Under lightweight constraints, adding only 0.9M parameters, GeoMamba achieves 86.57% classification accuracy on the most challenging ScanObjectNN subset PB-T50-RS without pretraining, surpassing PointMamba by 4.09 percentage points. With pretrained weights, the accuracy further improves to 89.21%. Through extensive ablation experiments and visualization analysis, we validate the effectiveness and interpretability of the dynamic step-size modulation mechanism.