Research map / Detection and segmentation / Open-set and few-shot detection
Addressing Background Context Bias in Few-Shot Segmentation through Iterative Modulation
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 an iterative few-shot segmentation network (ABCB) that modulates support foreground features with query background context and cleanses noise via confidence-biased attention to counter background context bias.
Problem
Few-shot segmentation methods use support foreground prototypes that become misaligned with query foreground features when support and query backgrounds differ (background context bias).
Approach
Iterates three steps (query prediction, support modulation using query background context derived from an input-to-output evolution feature, and information cleansing via confidence-biased attention) so that support prototypes are refined toward query foreground features.
Key results
- On PASCAL-5i with ResNet101, achieves 72.0% 1-shot and 74.9% 5-shot mIoU, outperforming the second-best method by 1.9% and 2.7%.
- On COCO-20i with ResNet101, achieves 51.5% 1-shot and 58.8% 5-shot mIoU, beating the second-best method by 1.7% and 1.8%.
- Ablation on PASCAL-5i fold-0 (ResNet50): baseline 61.48% mIoU rises to 72.92% with support modulation and information cleansing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ABCB instead of MSI | ResNet101 · PASCAL-5i 1-shot mean | mIoU: 70.1 -> 72.0 | Helped |
| use ABCB instead of SCCAN | ResNet101 · PASCAL-5i 5-shot mean | mIoU: 71.5 -> 74.9 | Helped |
| use ABCB instead of MSI | ResNet101 · COCO-20i 1-shot mean | mIoU: 49.8 -> 51.5 | Helped |
| use ABCB instead of SCCAN | ResNet101 · COCO-20i 5-shot mean | mIoU: 57.0 -> 58.8 | Helped |
| add Support Modulation (SM) (ablation) | Backbone + QP baseline · PASCAL-5i fold-0 | mIoU: 61.48 -> 68.20 | Helped |
| add Information Cleansing (IC) (ablation) | Backbone + QP + SM · PASCAL-5i fold-0 | mIoU: 68.20 -> 72.92 | Helped |
| add pixel-wise evolution feature Ep (ablation) | ABCB support modulation · PASCAL-5i fold-0 | mIoU: 68.85 -> 72.92 | Helped |
| add structure-wise evolution feature Es (ablation) | ABCB support modulation · PASCAL-5i fold-0 | mIoU: 70.29 -> 72.92 | Helped |
| increase iterations from T=1 to T=3 (ablation) | ABCB · PASCAL-5i fold-0 | mIoU: 61.48 -> 72.92 | Helped |
| increase iterations from T=3 to T=5 (ablation) | ABCB · PASCAL-5i fold-0 | mIoU: 72.92 -> 73.05 | No clear effect |
| use prediction from later iterations (ablation) | ABCB · PASCAL-5i fold-0 | mIoU: 63.08 -> 70.76 -> 72.92 | Helped |
Surprising findings
- Iterations beyond T=3 give only marginal gains (72.92 to 73.05 at T=5) while increasing MACs.
Limitations
- Computation and parameter count increase over a single-step baseline.
- Evaluated only on PASCAL-5i and COCO-20i with ResNet backbones.
Experimental setup
Datasets: PASCAL-5i COCO-20i
Benchmarks: PASCAL-5i COCO-20i
Metrics: mIoU
Baselines: NTRENet BAM AAFormer SSP IPMT ABCNet HDMNet MIANet MSI SCCAN DCAMA VAT
Builds on: BAM HDMNet IPMT
Closest papers by meaning
- Task-Disruptive Background Suppression for Few-Shot Segmentation (AAAI 2024 · similarity 0.91)
- Foreground-Covering Prototype Generation and Matching for SAM-Aided Few-Shot Segmentation (AAAI 2025 · similarity 0.88)
- Bidirectional Reciprocative Information Communication for Few-Shot Semantic Segmentation (ICML 2024 · similarity 0.85)
- Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.85)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.84)
- Hybrid Mamba for Few-Shot Segmentation (NeurIPS 2024 · similarity 0.84)
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.