Research map / Detection and segmentation / Open-set and few-shot detection
Hybrid Mamba for Few-Shot Segmentation
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 HMNet, a hybrid Mamba network for few-shot segmentation that uses support-recapped and query-intercepted Mamba blocks to fuse support foreground into query features with linear complexity.
Problem
Few-shot segmentation needs to fuse support foreground into query features, but cross attention is quadratic in pixel count, and naive cross Mamba suffers from support forgetting and intra-class gap issues.
Approach
Adapts Mamba into a cross-attention-like module: a support recapped Mamba periodically re-scans downsampled support features interleaved with query patches so the hidden state retains support information, and a query intercepted Mamba processes query pixels in parallel from the averaged support hidden state so query pixels cannot attend to each other and must use support features.
Key results
- On PASCAL-5i with ResNet50 1-shot, HMNet reaches 70.4 mIoU, versus 69.4 for HDMNet and 69.0 for SVF.
- On COCO-20i with VGG16 1-shot, HMNet reaches 49.1 mIoU, a 3.2 point gain over HDMNet (45.9).
- In the component ablation, adding SRM and QIM (shared parameters) raises the basic Mamba model from 67.4 to 68.6 mIoU, exceeding the attention-based SCCAN-style baseline at 66.8 under the same condition.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest baseline with HMNet | ResNet50 few-shot segmentation · PASCAL-5ᵢ 1-shot | mIoU: 70.1 -> 70.4 | Helped |
| replace the strongest baseline with HMNet | ResNet50 few-shot segmentation · PASCAL-5ᵢ 5-shot | mIoU: 73.2 -> 74.1 | Helped |
| replace the strongest baseline with HMNet | ResNet50 few-shot segmentation · COCO-20ᵢ 1-shot | mIoU: 51.0 -> 52.1 | Helped |
| replace the strongest baseline with HMNet | ResNet50 few-shot segmentation · COCO-20ᵢ 5-shot | mIoU: 57.3 -> 58.9 | Helped |
| add self Mamba blocks to the basic model (ablation) | Basic few-shot segmentation model · PASCAL-5ᵢ 1-shot | mIoU: 65.2 -> 67.4 | Helped |
| replace sequential support-then-query scanning with periodic support recap (ablation) | Basic SRM · PASCAL-5ᵢ 1-shot | mIoU: 67.5 -> 68.1 | Helped |
| share QIM parameters with the first SRM (ablation) | Query intercepted Mamba · PASCAL-5ᵢ 1-shot | mIoU: 67.8 -> 68.1 | Helped |
| combine support recap and query interception (ablation) | Hybrid Mamba block · PASCAL-5ᵢ 1-shot | mIoU: 68.1 -> 68.6 | Helped |
| add BAM finetuned backbones and base-class annotations (ablation) | HMNet · PASCAL-5ᵢ 1-shot | mIoU: 68.6 -> 70.4 | Helped |
| increase the number of Mamba blocks from 4 to 6 (ablation) | HMNet · PASCAL-5ᵢ 1-shot | mIoU: 67.8 -> 68.2 | Helped |
| increase the number of Mamba blocks from 6 to 8 (ablation) | HMNet · PASCAL-5ᵢ 1-shot | mIoU: 68.2 -> 68.6 | Helped |
| increase the number of Mamba blocks from 8 to 10 (ablation) | HMNet · PASCAL-5ᵢ 1-shot | mIoU: 68.6 -> 67.5 | Hurt |
Surprising findings
- The plain sequential cross Mamba gives only a 0.1 mIoU gain over self Mamba alone, which the authors attribute to support forgetting.
- Sharing Mamba parameters between the first SRM and QIM improves mIoU by 0.3 points, despite QIM's single-step scan.
Limitations
- QIM has not yet been implemented with CUDA, making training and testing slower.
- Results rely on the base-class ensemble from BAM for the best scores (70.4 vs 68.6 without it).
Experimental setup
Datasets: PASCAL-5i COCO-20i PASCAL VOC 2012 SBD MSCOCO
Benchmarks: PASCAL-5i COCO-20i
Metrics: mIoU FB-IoU
Baselines: PFENet DACM FECANet SCCAN BAM SVF HDMNet CyCTR VAT ABCNet AMNet
Builds on: Mamba VMamba BAM SCCAN
Closest papers by meaning
- Unlocking the Power of SAM 2 for Few-Shot Segmentation (ICML 2025 · similarity 0.86)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.86)
- Foreground-Covering Prototype Generation and Matching for SAM-Aided Few-Shot Segmentation (AAAI 2025 · similarity 0.86)
- Adaptive FSS: A Novel Few-Shot Segmentation Framework via Prototype Enhancement (AAAI 2024 · similarity 0.86)
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.85)
- Task-Disruptive Background Suppression for Few-Shot Segmentation (AAAI 2024 · similarity 0.85)
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.