Research map / Detection and segmentation / CNN features
DAMamba: Vision State Space Model with Dynamic Adaptive Scan
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAMamba, a vision state space model whose Dynamic Adaptive Scan (DAS) learns offsets to adaptively sample and order patches, achieving SOTA-level ImageNet, COCO, and ADE20K results among vision SSMs.
Problem
Existing vision SSMs (Vim, VMamba, PlainMamba, LocalMamba) rely on manually designed, input-independent scanning orders that break 2D spatial adjacency and lack flexibility to capture complex image structures.
Approach
Dynamic Adaptive Scan (DAS) uses an offset prediction network to predict per-patch 2D sampling offsets, samples features via bilinear interpolation at the shifted positions, and re-orders them top-to-bottom, left-to-right before feeding them into the SSM, so the scan adapts to each input image.
Key results
- DAMamba-T reaches 83.8% ImageNet-1K top-1 at 26M params and 4.8G FLOPs, versus 82.6% for VMamba-T; DAMamba-B reaches 85.2%.
- With Mask R-CNN 1x, DAMamba-T/S/B reach 48.5/49.8/50.6 box AP, outperforming VMamba by 1.2/1.1/1.4 AP.
- On ADE20K with UPerNet (SS), DAMamba-T/S/B reach 50.3/51.2/51.9 mIoU, versus 48.0 for VMamba-T.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dynamic adaptive scanning (DAS) instead of sweeping scan (ablation) | DAMamba-T · ImageNet-1K | Top-1 accuracy (%): 80.8 -> 81.2; 0.4% | Helped |
| Add convolutional positional encoding (Convpos) (ablation) | DAMamba-T with DAScan · ImageNet-1K | Top-1 accuracy (%): 81.2 -> 81.4; 0.2% | Helped |
| Add convolutional FFN (ConvFFN) (ablation) | DAMamba-T with DAScan and Convpos · ImageNet-1K | Top-1 accuracy (%): 81.4 -> 81.8; 0.4% | Helped |
| Use exponential moving average (EMA) during training | DAMamba · ImageNet-1K | Final model performance | No clear effect |
| Use DAMamba-T instead of VMamba-T as backbone | Mask R-CNN · COCO2017 val, 1× schedule | APb: 47.3 -> 48.5 | Helped |
| Use DAMamba-T instead of VMamba-T as backbone | Mask R-CNN · COCO2017 val, 1× schedule | APm: 42.7 -> 43.4 | Helped |
| Use DAMamba-T instead of VMamba-T as backbone | UPerNet · ADE20K, single-scale testing | mIoU: 48.0 -> 50.3 | Helped |
| Use DAMamba-T instead of VMamba-T as backbone | Image classification model · ImageNet-1K val | Top-1 accuracy (%): 82.6 -> 83.8 | Helped |
Limitations
- Ablation gains are modest (DAS adds 0.4% top-1 on DAMamba-T), and the ablation is only on ImageNet classification.
- Visualization is limited to three ImageNet examples and requires mapping floating-point positions to nearest patches.
Experimental setup
Datasets: ImageNet-1K COCO 2017 ADE20K
Benchmarks: ImageNet-1K COCO 2017 ADE20K
Metrics: top-1 accuracy AP^b AP^m mIoU FLOPs throughput
Baselines: VMamba LocalVMamba Vim ConvNeXt Swin Transformer UniRepLKNet TransNeXt MambaOut BiFormer NAT
Builds on: Mamba VMamba Vim S4 DeiT
Closest papers by meaning
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.89)
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.87)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.87)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · 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.