Research map / Architectures and efficiency / CNN design
EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba
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 EfficientVMamba, a lightweight visual state space network that uses atrous (skip-sampled) selective scanning, a parallel convolution branch with SE fusion, and inverted placement of SSM and convolution blocks across stages.
Problem
Lightweight vision models face a trade-off between accuracy and efficiency: CNNs lose global context while Transformers' global self-attention scales as O(N^2); existing SSM vision backbones like VMamba remain computationally heavy for resource-constrained use.
Approach
An efficient 2D selective scan (ES2D) that skip-samples patches with step p and scans each sub-grid, reducing tokens per scan from N to N/p^2 while keeping a global receptive field; combined with a parallel 3x3 conv branch fused via SE in an EVSS block, and an inverted stage layout that puts SSM blocks in early high-resolution stages and MobileNetV2 blocks in deeper stages.
Key results
- EfficientVMamba-T reaches 76.5% ImageNet top-1 at 0.8 GFLOPs, versus 72.1% for VMamba-T in the matched-FLOPs ablation.
- EfficientVMamba-S reaches 78.7% top-1 at 1.3 GFLOPs, improving Vim-Ti (73.1% at 1.5 GFLOPs) by 5.6 points.
- EfficientVMamba-B reaches 81.8% top-1 at 4.0 GFLOPs, versus VMamba-T's 82.2% at 5.6 GFLOPs with 22M params.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EfficientVMamba-T instead of RegNetY-800M | ImageNet-1K classification · ImageNet-1K | Top-1 accuracy: 76.3 -> 76.5 | Helped |
| Use EfficientVMamba-S instead of MobileViT-S | ImageNet-1K classification · ImageNet-1K | Top-1 accuracy: 78.4 -> 78.7 | Helped |
| Use EfficientVMamba-B instead of VMamba-T | ImageNet-1K classification · ImageNet-1K | Top-1 accuracy: 82.2 -> 81.8 | Hurt |
| Use EfficientVMamba-T instead of PVTv2-b0 | RetinaNet backbone · COCO 2017 | AP: 37.2 -> 37.5 | Helped |
| Use EfficientVMamba-S instead of PVTv2-b1 | RetinaNet backbone · COCO 2017 | AP: 41.2 -> 39.1 | Hurt |
| Use EfficientVMamba-B instead of PVTv1-Medium | RetinaNet backbone · COCO 2017 | AP: 41.9 -> 42.8 | Helped |
| Use EfficientVMamba-B instead of VMamba-S | UperNet backbone · ADE20K val, single-scale | mIoU: 49.5 -> 46.5 | Hurt |
| Add ES2D by replacing SS2D (ablation) | EfficientVMamba-T · ImageNet | accuracy: 72.1 -> 73.6 | Helped |
| Add ES2D by replacing SS2D (ablation) | EfficientVMamba-B · ImageNet | accuracy: 80.2 -> 80.9 | Helped |
| Add the convolutional Fusion branch with SE (ablation) | EfficientVMamba-T with ES2D · ImageNet | accuracy: 73.6 -> 75.1 | Helped |
| Add the convolutional Fusion branch with SE (ablation) | EfficientVMamba-B with ES2D · ImageNet | accuracy: 80.9 -> 81.2 | Helped |
| Add InRes in stages 3–4 while using EVSS in stages 1–2 (ablation) | EfficientVMamba-T · ImageNet | accuracy: 75.6 -> 76.5 | Helped |
Surprising findings
- Placing convolution (InRes) blocks in the deep stages works better than placing them early, the opposite of the layout used in lightweight ViTs.
- The ES2D skip-sampling scan slightly raised accuracy over the full SS2D scan at the base scale (80.2% to 80.9%) while lowering FLOPs.
Limitations
- SSM computation is more intricate and harder to parallelize efficiently, per the authors' limitations section.
- Comparisons are against a VMamba baseline with adjusted dimensions to match FLOPs, and several cross-paper comparisons mix differing training recipes and detection settings.
Experimental setup
Datasets: ImageNet-1K MSCOCO 2017 ADE20K
Benchmarks: ImageNet-1K MSCOCO 2017 ADE20K
Metrics: Top-1 accuracy AP mIoU Params FLOPs
Baselines: VMamba-T Vim-Ti MobileViT-S PVTv2-B0 DeiT-Ti Swin-T ResNet-50 LightViT-T
Builds on: Mamba VMamba S4 MobileNetV2 Squeeze-and-Excitation
Closest papers by meaning
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.92)
- VMamba: Visual State Space Model (NeurIPS 2024 · similarity 0.91)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.90)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.90)
- QuadMamba: Learning Quadtree-based Selective Scan for Visual State Space Model (NeurIPS 2024 · similarity 0.90)
- EfficientViM: Efficient Vision Mamba with Hidden State Mixer based State Space Duality (CVPR 2025 · similarity 0.89)
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.