Research map / Detection and segmentation / Open-set and few-shot detection
PR-MaGIC: Prompt Refinement Via Mask Decoder Gradient Flow For In-Context Segmentation
Minjae Lee, Sungwoo Hur, Soojin Hwang, Won Hwa Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PR-MaGIC gradient-flow refinement with top-1 selection | PerSAM-F · FSS-1000 | mIoU: 85.85 -> 87.26 | Improved |
| add PR-MaGIC with top-1 selection | PerSAM-F · COCO-20i | mIoU: 53.36 -> 55.96 | Improved |
| add PR-MaGIC with top-1 selection | Matcher (1-shot) · FSS-1000 | mIoU: 92.08 -> 92.06 | No clear effect |
| add PR-MaGIC with top-1 selection | Matcher (1-shot) · DIS5K | mIoU: 46.65 -> 55.08 | Improved |
| add PR-MaGIC with top-1 selection | Matcher (5-shot) · COCO-20i | mIoU: 67.69 -> 70.74 | Improved |
| add PR-MaGIC with top-1 selection | Matcher (1-shot) · PACO-Part | mIoU: 50.27 -> 54.08 | Improved |
| use oracle selection instead of top-1 similarity selection (ablation) | PR-MaGIC on PerSAM-F · FSS-1000 | mIoU: 87.26 -> 91.88 | Improved |
| use a fixed number of refinement iterations instead of top-1 selection (ablation) | PR-MaGIC on Matcher (1-shot) · COCO-20i | mIoU gain: -0.15, 0.21, 0.07, 0.31, 0.66 | Mixed |
| use fixed 5 iterations instead of top-1 selection (ablation) | PR-MaGIC on PerSAM-F (1-shot) · COCO-20i | mIoU gain: -0.43 vs +2.60 (Top-1) | Hurt |
| increase step size eta to 1e-2 (ablation) | PR-MaGIC · COCO-20i (100 train samples) | mIoU: rapid early gains but degrades with more iterations | Mixed |
| decrease step size eta to 1e-4 or 1e-5 (ablation) | PR-MaGIC · COCO-20i (100 train samples) | mIoU: saturates below best achievable mIoU | Hurt |
| larger step size eta (ablation) | PR-MaGIC on PerSAM-F · FSS-1000 (9 classes) | mIoU: 70.88 > 67.89 > 64.78 | Improved |
Closest papers by meaning
- Segment Anyword: Mask Prompt Inversion for Open-Set Grounded Segmentation (ICML 2025 · similarity 0.88)
- A Simple Image Segmentation Framework via In-Context Examples (NeurIPS 2024 · similarity 0.88)
- Stable Segment Anything Model (ICLR 2025 · similarity 0.88)
- Visual In-Context Prompting (CVPR 2024 · similarity 0.87)
- Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts (CVPR 2024 · similarity 0.87)
- AoP-SAM: Automation of Prompts for Efficient Segmentation (AAAI 2025 · 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.
