Research map / Architectures and efficiency / Attention and transformers
PVMamba: Parallelizing Vision Mamba via Dynamic State Aggregation
Fei Xie, Zhongdao Wang, Weijia Zhang, Chao Ma
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 |
|---|---|---|---|
| replace SS2D with Sa-DSA operator (ablation) | Mamba baseline · ImageNet-1K | Top-1 accuracy (%): 82.5 -> 83.0 | Improved |
| add hybrid model stage design (ablation) | PVMamba tiny · ImageNet-1K | Top-1 accuracy (%): 83.0 -> 83.2 | Improved |
| add MESA (ablation) | PVMamba tiny · ImageNet-1K | Top-1 accuracy (%): 83.5 -> 83.9 | Improved |
| use Lo-DSA in the first two stages instead of Sa-DSA (ablation) | PVMamba lite · ImageNet-1K | Top-1 accuracy (%): 82.0 -> 82.2 | Improved |
| use Lo-DSA in first two stages instead of Sa-DSA (ablation) | PVMamba lite · ImageNet-1K | Throughput (img/s): 1128 -> 1423 | Improved |
| replace stage 4 DSA with vanilla attention (ablation) | PVMamba lite · ImageNet-1K | Top-1 accuracy (%): 82.2 -> 82.2 | No clear effect |
| replace stage 4 DSA with vanilla attention (ablation) | PVMamba lite · ImageNet-1K | Throughput (img/s): 1423 -> 1524 | Improved |
| increase sampling points Kd from 3 to 5 (ablation) | PVMamba lite · ImageNet-1K | Top-1 accuracy (%): 82.1 -> 82.2 | Mixed |
| increase sampling points Kd from 3 to 9 (ablation) | PVMamba lite · COCO | box AP: 45.2 -> 45.7 | Improved |
| increase sampling points Kd from 5 to 9 (ablation) | PVMamba lite · ImageNet-1K | Memory cost: 5.8GB -> 6.5GB | Hurt |
| add layer multiplexing in interleaved pattern B (ablation) | Sa-DSA stack · ImageNet-1K | training stability | Improved |
| replace sequential SSM with parallel DSA, no acceleration (ablation) | SSM · latency | Latency (ms): 0.613 -> 0.702 | Mixed |
Closest papers by meaning
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.90)
- VMamba: Visual State Space Model (NeurIPS 2024 · similarity 0.89)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.89)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.89)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.88)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.88)
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.
