Research map / Detection and segmentation / CNN features
SF-Mamba: Rethinking State Space Model for Vision
Masakazu Yoshimura, Teruaki Hayashi, Yuki Hoshino, Wei-Yao Wang, Takeshi Ohashi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add auxiliary token swapping (ablation) | MambaVision-T · ImageNet-1K / ADE20K | top-1 / mIoU: 82.2 -> 82.5 (IN1K); 46.0 -> 47.2 (ADE20K) | Improved |
| add learnable auxiliary tokens without swapping (ablation) | MambaVision-T · ImageNet-1K / ADE20K | top-1 / mIoU: 82.2 -> 82.1 (IN1K); 46.0 -> 46.2 (ADE20K) | Mixed |
| initialize aux tokens as global average instead of learnable (ablation) | SF-Mamba-T · ImageNet-1K / ADE20K | top-1 / mIoU: 82.3 -> 82.4 (IN1K); 46.5 -> 46.8 (ADE20K) | Improved |
| discard aux token after first attention layer instead of before attention (ablation) | SF-Mamba-T · ADE20K | mIoU: 46.8 -> 47.2 | Improved |
| discard aux token after all attention instead of after first attention (ablation) | SF-Mamba-T · ADE20K | mIoU: 47.2 -> 46.6 | Hurt |
| add batch folding (BFold) with B1=1 (ablation) | MambaVision-T · ImageNet-1K throughput | img/s: 6989 -> 7601 | Improved |
| use adaptive B1 with folding-compatible 1-D conv (ablation) | MambaVision-T · ImageNet-1K throughput | img/s: 7279 -> 7685 | Improved |
| adaptive B1 without compatible conv vs fixed B1=4 (ablation) | MambaVision-T · ImageNet-1K throughput | img/s: 7641 -> 7279 | Hurt |
| batch folding of SSM kernel at short sequences (ablation) | Mamba SSM kernel · SSM kernel speed | speedup: 110% to 180% | Improved |
| replace uni-scan with parallel bi-scan (add) (ablation) | MambaVision-T · ImageNet-1K | top-1 / img/s: 82.2 -> 82.3; 6979 -> 6235 img/s | Mixed |
| replace uni-scan with series bi-scan (ablation) | MambaVision-T · ImageNet-1K | top-1: 82.2 -> 82.3 | No clear effect |
| replace uni-scan with swap on Mamba-only macro-architecture (ablation) | MambaVision-T w/o Attention · ImageNet-1K | top-1: 80.2 -> 81.0 | Improved |
Closest papers by meaning
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.92)
- QuadMamba: Learning Quadtree-based Selective Scan for Visual State Space Model (NeurIPS 2024 · similarity 0.92)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.91)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.91)
- Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model (ICML 2024 · similarity 0.91)
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
