Research map / Detection and segmentation / Open-set and few-shot detection
SAM-Aware Graph Prompt Reasoning Network for Cross-Domain 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 GPRN, which turns SAM masks into semantic visual prompts, reasons over them with a graph attention network, and refines predictions by feeding SAM point prompts selected from the initial prediction, reaching SOTA on four CD-FSS benchmarks.
Problem
Cross-domain few-shot segmentation (CD-FSS) suffers from domain shift between training and test data; existing visual prompt methods lack semantic and spatial information and are tied to transformers.
Approach
Use SAM-generated masks to initialize semantic-aware visual prompts (masked average pooling), connect them in a graph attention network to enforce consistency across over-segmented sub-regions, map them back to spatial feature maps via reverse MAP, and at test time refine predictions with an adaptive, parameter-free point-selection scheme that prompts SAM.
Key results
- GPRN reaches 71.7% average mean-IoU in 1-shot and 75.3% in 5-shot, versus 67.8% and 71.9% for the best prior fine-tuning method (IFA/DARNet).
- Ablation on FSS-1000: SPI adds 4.1 points over the 73.8% baseline, and the full model (SPI+GPR+APS) reaches 81.1% 1-shot mean-IoU.
- Compared with PerSAM and APSeg in 1-shot, GPRN improves by 34.1 and 10.4 points respectively, per the paper's text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPRN instead of IFA | CD-FSS with ResNet-50 · Average 1-shot | mean-IoU: 67.8 -> 71.7 | Helped |
| Use GPRN instead of IFA | CD-FSS with ResNet-50 · Average 5-shot | mean-IoU: 71.4 -> 75.3 | Helped |
| Add SPI to the baseline (ablation) | GPRN baseline · FSS-1000 1-shot | mean-IoU: 73.8 -> 77.9 | Helped |
| Add GPR after SPI (ablation) | GPRN with SPI · FSS-1000 1-shot | mean-IoU: 77.9 -> 78.4 | Helped |
| Add APS to the baseline (ablation) | GPRN baseline · FSS-1000 1-shot | mean-IoU: 73.8 -> 77.2 | Helped |
| Add SPI, GPR, and APS (ablation) | GPRN baseline · FSS-1000 1-shot | mean-IoU: 73.8 -> 81.1 | Helped |
| Increase alpha from 0.1 to 0.3 (ablation) | GPRN · FSS-1000 1-shot | mean-IoU: 81.1 -> 80.9 | Hurt |
| Increase alpha from 0.1 to 0.5 (ablation) | GPRN · FSS-1000 1-shot | mean-IoU: 81.1 -> 80.5 | Hurt |
| Increase mask count from 40 to 60 (ablation) | GPRN · FSS-1000 1-shot | mean-IoU: 81.1 -> 81.3 | No clear effect |
| Add negative point prompts to 10 positive prompts (ablation) | GPRN APS · FSS-1000 1-shot | mean-IoU: 79.2 -> 80.5 | Helped |
| Use APS point selection instead of PerSAM selection (ablation) | GPRN APS · FSS-1000 1-shot | mean-IoU: 77.9 -> 81.1 | Helped |
| Add box prompts to APS (ablation) | GPRN APS · FSS-1000 1-shot | mean-IoU: 81.1 -> 81.1 | No clear effect |
Surprising findings
- Training on PASCAL VOC gives no gain over the no-training setting, while the target-domain fine-tuning phase adds about 4 points.
Limitations
- SAM is run on upsampled 1024x1024 inputs and generates up to 40 masks per image, which adds inference cost.
- Test-time refinement depends on the quality of the initial prediction and the fixed fusion weight beta.
Experimental setup
Datasets: PASCAL VOC SDS Deepglobe ISIC Chest X-Ray FSS-1000
Benchmarks: Deepglobe ISIC Chest X-Ray FSS-1000
Metrics: mean-IoU
Baselines: AMP RestNet PerSAM ABCDFSS PMNet DRANet APSeg PATNet DARNet IFA DMTNet
Builds on: SSP PATNet Segment Anything Model graph attention network PerSAM maskSLIC
Closest papers by meaning
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · similarity 0.91)
- APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation (CVPR 2024 · similarity 0.90)
- Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach (CVPR 2024 · similarity 0.89)
- DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation (CVPR 2025 · similarity 0.88)
- AoP-SAM: Automation of Prompts for Efficient Segmentation (AAAI 2025 · similarity 0.88)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.87)
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.