Research map / Detection and segmentation / Open-set and few-shot detection
Task-Disruptive Background Suppression 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 Task-disruptive Background Suppression (TBS), a plug-in module that down-weights harmful support-background features in affinity-learning few-shot segmentation using query-relevant and target-relevant scores, improving DCAMA and CyCTR on PASCAL-5i and COCO-20i.
Problem
Few-shot segmentation methods that use the entire support background for affinity learning can misclassify query foreground as background when support background is dissimilar to query background or contains objects similar to the query target.
Approach
Two spatial-wise scores—a query-relevant score (how well support background is reconstructed from query features via cross-attention) minus a target-relevant score (how well it is reconstructed from support foreground)—form a query background-relevant score that scales support background features to suppress disruptive ones.
Key results
- DCAMA + TBS with Swin-B reaches 71.2 mIoU 1-shot on PASCAL-5i versus 69.3 for DCAMA (Swin-B).
- CyCTR + TBS with ResNet-101 reaches 64.0 mIoU 1-shot on PASCAL-5i versus 63.7 for CyCTR.
- DCAMA + TBS with Swin-B reaches 52.3 mIoU 1-shot on COCO-20i versus 50.9 for DCAMA (Swin-B).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TBS to DCAMA (ResNet101) | DCAMA · PASCAL-5ᵢ | mIoU: 64.6/68.3 -> 65.9/70.5 | Helped |
| Add TBS to CyCTR (ResNet101) | CyCTR · PASCAL-5ᵢ | mIoU: 63.7/67.4 -> 64.0/68.0 | Helped |
| Add TBS to DCAMA (Swin-B) | DCAMA · PASCAL-5ᵢ | mIoU: 69.3/74.9 -> 71.2/75.2 | Helped |
| Add TBS to DCAMA (Swin-B) | DCAMA · COCO-20ᵢ | mIoU: 50.9/58.3 -> 52.3/58.5 | Helped |
| Add query-relevant scoring without target-relevant scoring (ablation) | DCAMA + TBS (Swin-B) · PASCAL-5ᵢ fold-1, 1-shot | mIoU: 73.8 -> 73.4 | Hurt |
| Add target-relevant scoring without query-relevant scoring (ablation) | DCAMA + TBS (Swin-B) · PASCAL-5ᵢ fold-1, 1-shot | mIoU: 73.8 -> 74.1 | Helped |
| Add query-relevant scoring alongside target-relevant scoring (ablation) | DCAMA + TBS (Swin-B) · PASCAL-5ᵢ fold-1, 1-shot | mIoU: 74.1 -> 74.4 | Helped |
Surprising findings
- The query-relevant score alone slightly reduces mIoU (73.8 to 73.4), and only the combination with the target-relevant score improves performance.
- Background attention scores decrease with TBS even though foreground-pair attention increases, which the authors attribute to unintended suppression of useful background.
Limitations
- TBS is applied only at some scales for DCAMA with ResNet-101 due to memory limitations.
- Background attention scores decrease under TBS, suggesting unintended suppression of useful background features.
- Gains on COCO-20i 5-shot are less pronounced than in 1-shot.
Experimental setup
Datasets: PASCAL-5i COCO-20i
Benchmarks: PASCAL-5i COCO-20i
Metrics: mIoU FB-IoU
Baselines: CWT DoG-LSTM DAN HSNet CyCTR DCAMA
Builds on: DCAMA CyCTR HSNet
Closest papers by meaning
- Addressing Background Context Bias in Few-Shot Segmentation through Iterative Modulation (CVPR 2024 · similarity 0.91)
- Foreground-Covering Prototype Generation and Matching for SAM-Aided Few-Shot Segmentation (AAAI 2025 · similarity 0.89)
- Unlocking the Power of SAM 2 for Few-Shot Segmentation (ICML 2025 · similarity 0.85)
- Hybrid Mamba for Few-Shot Segmentation (NeurIPS 2024 · similarity 0.85)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.85)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 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.