LBMamba: Locally Bi-directional Mamba

The recent introduction of LBMamba, a locally bi-directional State Space Model (SSM), represents a pivotal development in the field of artificial intelligence, particularly in computer vision. Traditional Mamba models, while efficient, faced limitations due to their unidirectional nature, which restricted their ability to access future states. LBMamba addresses this by integrating a lightweight backward scan into the forward scan, effectively maintaining computational efficiency without the burden of additional scans. This innovation is further exemplified in LBVim, a backbone that alternates scan directions every two layers, achieving notable improvements in accuracy across various datasets. For instance, LBVim demonstrates a 1.2% increase in top-1 accuracy on ImageNet-1K, a 1.65% improvement in mean Intersection over Union (mIoU) on ADE20K, and enhancements in detection metrics on COCO. These advancements not only highlight the superior performance-throughput trade-off offered by LBM…