PointMamba

Why we included this project

PointMamba is a NeurIPS 2024 implementation of a state space model for point cloud analysis. It replaces the quadratic-complexity attention of standard transformers with a linear-complexity Mamba backbone, which makes it a useful baseline for global modeling on unordered point sets when self-attention's memory cost is a problem. The repo includes pretraining and fine-tuning configs plus checkpoints for ModelNet40 and ScanObjectNN, so reproducing the reported numbers or adapting the encoder to your own task is fairly direct. Because it follows the Point-BERT and Point-MAE conventions, comparing it against those transformer-based predecessors is easy. Teams exploring efficient attention alternatives in 3D vision, and students benchmarking state space models, will get the most value here.