SF-Mamba: Rethinking State Space Model for Vision

TL;DR

SF-Mamba introduces a new vision model that enhances the efficiency and effectiveness of the Mamba architecture by enabling bidirectional information flow and better GPU parallelism, outperforming existing methods across multiple vision tasks.

Contribution

The paper proposes SF-Mamba, a novel vision model that improves upon Mamba by incorporating auxiliary patch swapping and batch folding, addressing limitations in non-causal interactions and computational speed.

Findings

Outperforms state-of-the-art baselines in image classification, detection, and segmentation.

Achieves higher throughput across various model sizes.

Demonstrates significant efficiency and accuracy improvements.

Abstract

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.