Research map / Detection and segmentation / Open-set and few-shot detection
Bidirectional Reciprocative Information Communication for Few-Shot Semantic 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 IFRNet, which iteratively passes category information bidirectionally between query and support (via a Query Feedback Branch, Query Amplifier Branch, and Query Rectification Module) to address intra-class diversity in few-shot semantic segmentation.
Problem
Few-shot semantic segmentation methods use one-way support-to-query category flow, ignoring intra-class diversity between support and query objects.
Approach
Introduces cybernetics-inspired information feedback: query information is fed back to the support to build a query-related support prototype, which then reactivates query targets; prediction disparity before and after support activation identifies hard positive/negative regions for query self-rectification, stacked iteratively in FRLs.
Key results
- 1-shot PASCAL-5i mIoU of 67.8% (VGG-16) and 70.7% (ResNet-50), exceeding prior SOTA by 6.1% and 2.4%.
- 5-shot PASCAL-5i mIoU of 69.8% (VGG-16) and 72.1% (ResNet-50), exceeding prior SOTA by 4.5% and 1.0%.
- Ablation: QFB adds 3.1% mIoU over the baseline; full model (QFB+QAB+QRM) reaches 72.1% mIoU on PASCAL-5i 1-shot with ResNet-50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IFRNet instead of MSI | ResNet-50 few-shot segmentation · PASCAL-5i 1-shot | mIoU: 68.3 -> 70.7 | Helped |
| Use IFRNet instead of MSI | ResNet-50 few-shot segmentation · PASCAL-5i 5-shot | mIoU: 71.1 -> 72.1 | Helped |
| Use IFRNet instead of MIANet | ResNet-50 few-shot segmentation · COCO-20i 1-shot | mIoU: 47.6 -> 47.0 | Hurt |
| Use IFRNet instead of SCCAN | ResNet-50 few-shot segmentation · COCO-20i 5-shot | mIoU: 52.3 -> 52.3 | No clear effect |
| Add QFB to the baseline (ablation) | IFRNet on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 66.8 -> 69.9; 3.1% mIoU improvement | Helped |
| Add QAB to QFB (ablation) | IFRNet on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 69.9 -> 70.8 | Helped |
| Add the supplement prototype (QRM-su) (ablation) | IFRNet with QFB and QAB on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 70.8 -> 71.5; 0.7% mIoU improvement | Helped |
| Add the exclusion prototype (QRM-ex) (ablation) | IFRNet with QFB and QAB on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 70.8 -> 71.6; 0.8% mIoU improvement | Helped |
| Add QRM-su and QRM-ex (ablation) | IFRNet with QFB and QAB on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 70.8 -> 72.1 | Helped |
| Use 2 FRL layers instead of 1 (ablation) | IFRNet on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 70.8 -> 72.1; 1.3% mIoU | Helped |
| Increase FRL layers from 2 to 3 (ablation) | IFRNet on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 72.1 -> 72.5 | Helped |
| Increase FRL layers from 2 to 4 (ablation) | IFRNet on PASCAL-5i · PASCAL-5i 1-shot | mIoU: 72.1 -> 72.8 | Helped |
Surprising findings
- Adding more than 2 FRL layers yields little gain (e.g., 72.1 at 2 layers vs 73.0 at 5), unlike typical stacked ViT architectures.
Limitations
- Stacking more than 2 FRL layers does not substantially improve performance, unlike typical ViT stacking.
- Relies on a benchmark-specific episodic training protocol on PASCAL-5i and COCO-20i only.
Experimental setup
Datasets: PASCAL-5i COCO-20i
Benchmarks: PASCAL-5i COCO-20i
Metrics: mIoU FB-IoU
Baselines: PFENet PMMs HSNet APANet NTRENet DPCN BAM FECANet MIANet InPNet MVPNet HDMNet
Builds on: PFENet BAM DCAMA IPMT
Closest papers by meaning
- Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.87)
- Unlocking the Potential of Pre-trained Vision Transformers for Few-Shot Semantic Segmentation through Relationship Descriptors (CVPR 2024 · similarity 0.86)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.86)
- DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation (CVPR 2025 · similarity 0.86)
- Enhancing Generalized Few-Shot Semantic Segmentation via Effective Knowledge Transfer (AAAI 2025 · similarity 0.86)
- A Surprisingly Simple Approach to Generalized Few-Shot Semantic Segmentation (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.