Density-Aware Multi-Level Geometry Enhancement for Mamba-Based Point Cloud Classification

Mamba-based point cloud networks process serialized point tokens with state space layers and offer efficient classification, yet their performance is largely determined by the quality of local tokens produced before serialization. Existing tokenization compresses local patches via symmetric pooling, thereby discarding fine-grained geometric relationships and introducing aggregation bias under non-uniform sampling. To handle this, we propose a density-aware multi-level geometry enhancement method. A density-weighted token encoder estimates the local point density within each patch and adaptively calibrates point-wise contributions before aggregation, thus reducing the dominance of redundant dense samples. A multi-level local geometry branch extracts hierarchical coordinate-based neighborhood features directly from raw points and injects them into serialized tokens through residual fusion, compensating for geometric details lost during patch compression. Supervised contrastive learning is further adopted as an auxiliary regularizer to improve intra-class compactness and inter-class separability. Experiments on ScanObjectNN and ModelNet40 confirm the effectiveness of the approach: our method achieves 94.42 ± 0.09%, 92.11 ± 0.13%, and 87.89 ± 0.11% on the OBJ_BG, OBJ_ONLY, and PB_T50_RS variants, respectively, while ModelNet40 accuracy reaches 93.04 ± 0.12%. These results, obtained with only 12.57 M parameters, indicate a favorable accuracy–efficiency trade-off on the evaluated benchmarks.