Point cloud learning plays a vital role in scenarios like robotic navigation, environmental perception, and autonomous driving. Point cloud analysis faces challenges, including quadratic computational complexity in Transformer-based models and reliance on static geometric serialization. The state space models also suffer from unidirectional modeling bias. To address these, this paper develops DyReMamba. It introduces a semantics-importance-based Dynamic Reordering Module and a Symmetric Bidirectional State Space Module. These modules enable semantics-driven adaptive sequence modeling and unbiased capture of global spatial dependencies. The Dynamic Reordering Module introduces a task-adaptive sequence reordering mechanism, leveraging multi-dimensional information from point clouds to enable the model to flexibly prioritize semantically critical features and break free from fixed geometric priors. The Symmetric Bidirectional State Space Module addresses the limitations of unidirectional modeling in traditional state space models, enabling the unbiased capture of global spatial dependencies in point clouds through bidirectional information exchange while maintaining efficient near-linear computational complexity. Experimental results show that DyReMamba demonstrates excellent classification performance on the ModelNet40 dataset and the OBJ-BG scenario of ScanObjectNN, achieving classification accuracy of 92.6% and 93.15% respectively. The mean IoU for all instances reaches 86.1% in the ShapeNetPart segmentation task. While maintaining near-linear computational complexity, it significantly improves accuracy for point cloud classification and segmentation tasks, creating a new paradigm for efficient point cloud analysis.
Li et al. (Sat,) studied this question.