Research map / Architectures and efficiency / Attention and transformers
Partial Ring Scan: Revisiting Scan Order in Vision State Space Models
Yi-Kuan Hsieh, Kuan-Chuan Peng, Xin Li, Ming-Ching Chang, Yu-Chee Tseng, Jun Wei Hsieh
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 partial channel filtering (PCF) to PRIS-Mamba (ablation) | PRIS-Mamba · ImageNet-1K val | Top-1 accuracy (%): 84.1 -> 84.5 | Improved |
| add PCF to PRIS-Mamba (ablation) | PRIS-Mamba · ImageNet-1K val | throughput (img/s): 2177 -> 2854 | Improved |
| plug PCF into VMamba (ablation) | VMamba · ImageNet-1K val | Top-1 accuracy (%): 82.6 -> 82.9 | Improved |
| plug PCF into other Vision-Mamba backbones (Vim, QuadMamba, Adventurer, SparX, GroupMamba) (ablation) | Vision-Mamba backbones · ImageNet-1K val | Top-1 accuracy (%): +0.2 to +0.4 pp | Improved |
| replace hard PCF with soft channel attention SE (ablation) | PRIS-Mamba w/o PCF · ImageNet-1K val | Top-1 accuracy (%): 84.1 -> 84.2; throughput 2177 -> 2089 | No clear effect |
| add CBAM soft channel attention instead of PCF (ablation) | PRIS-Mamba w/o PCF · ImageNet-1K val | Top-1 accuracy (%): 84.1 -> 84.3; params 27M -> 30M | No clear effect |
| use ring scan instead of fixed-path scans | PRIS-Mamba · ImageNet-1K val, 60 deg rotation | Top-1 accuracy (%): PRIS-Mamba 84.5 -> 84.4 vs VMamba 82.6 -> 80.6 | Improved |
| replace ring scan with single or multi-path fixed scans on the same backbone (ablation) | PRIS-Mamba-S · ImageNet-1K val | Top-1 accuracy (%): 84.9 vs 82.1-84.1 (best four-path S21 84.1) | Hurt |
| add more scan paths (1 to 2 to 4) to fixed-path scan (ablation) | fixed-path Vision Mamba · ImageNet-1K rotation 30/60 deg | Top-1 drop under rotation: losses of ~1.2-1.9 pp at 30/60 deg for all S1-S21 | No clear effect |
| vary ring width 1, 2, 4 (ablation) | PRIS-Mamba-T · ImageNet-1K val | Top-1 accuracy (%): 84.3 / 84.5 / 84.3 | No clear effect |
| use ring scan under random-mask occlusion | PRIS-Mamba 30M · ImageNet-1K val, mask 16x16 | Top-1 drop: -0.4 vs VMamba -0.9 | Improved |
| use ring scan under patch shuffling | PRIS-Mamba 30M · ImageNet-1K val, cut 16 patches | Top-1 drop: -0.7 vs VMamba -1.3 | Improved |
Closest papers by meaning
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.89)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.88)
- VMamba: Visual State Space Model (NeurIPS 2024 · similarity 0.87)
- REOrdering Patches Improves Vision Models (NeurIPS 2025 · similarity 0.87)
- Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · similarity 0.86)
- QuadMamba: Learning Quadtree-based Selective Scan for Visual State Space Model (NeurIPS 2024 · similarity 0.86)
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.
