Search papers, labs, and topics across Lattice.
The paper introduces SF-Mamba, a new visual Mamba architecture designed to overcome the limitations of unidirectional scanning in standard Mamba models for vision. SF-Mamba incorporates auxiliary patch swapping to enable bidirectional information flow and batch folding with periodic state reset to enhance GPU parallelism. Experiments across image classification, object detection, and segmentation tasks demonstrate that SF-Mamba achieves superior performance and throughput compared to existing state-of-the-art models.
SF-Mamba achieves state-of-the-art vision performance and improved throughput by rethinking Mamba's scan operation with bidirectional patch swapping and optimized GPU parallelism.
The realm of Mamba for vision has been advanced in recent years to strike for the alternatives of Vision Transformers (ViTs) that suffer from the quadratic complexity. While the recurrent scanning mechanism of Mamba offers computational efficiency, it inherently limits non-causal interactions between image patches. Prior works have attempted to address this limitation through various multi-scan strategies; however, these approaches suffer from inefficiencies due to suboptimal scan designs and frequent data rearrangement. Moreover, Mamba exhibits relatively slow computational speed under short token lengths, commonly used in visual tasks. In pursuit of a truly efficient vision encoder, we rethink the scan operation for vision and the computational efficiency of Mamba. To this end, we propose SF-Mamba, a novel visual Mamba with two key proposals: auxiliary patch swapping for encoding bidirectional information flow under an unidirectional scan and batch folding with periodic state reset for advanced GPU parallelism. Extensive experiments on image classification, object detection, and instance and semantic segmentation consistently demonstrate that our proposed SF-Mamba significantly outperforms state-of-the-art baselines while improving throughput across different model sizes. We will release the source code after publication.
