Reproduce “LocalMamba: Visual State Space Model with Windowed Selective Scan” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, Chang Xu
LocalMamba introduces a visual state space model that enhances the capture of 2D spatial dependencies through a windowed selective scan and an adaptive scan direction search. The model achieved 3.1% higher ImageNet-1K accuracy (76.2%) than Vim-Ti and improved object detection and semantic segmentation performance over existing Vision Mamba baselines.
Key insights
- Effectively modeling 2D local spatial dependencies is critical for Vision Mamba models to achieve competitive performance in image-based tasks.
- Optimal scanning patterns for Mamba architectures are layer-dependent, and adaptive selection significantly boosts performance.
- Combining diverse local and global scan directions with an attention mechanism for feature aggregation allows Vision Mamba to capture both fine-grained details and broader context efficiently.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.