Research map / Detection and segmentation / Open-set and few-shot detection
MFP: Making Full Use of Probability Maps for Interactive Image 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 MFP, which modulates previous-round probability maps with gamma correction around user clicks and feeds them, late-fused with backbone features, into click-based interactive segmentation networks trained with recursive click sampling.
Problem
Previous-round probability maps are used as network input in click-based interactive segmentation, but their useful information (object shape) is not well propagated to current predictions.
Approach
Modulate the previous probability map with click-dependent gamma correction (increase probabilities near foreground clicks, decrease near background clicks), feed the modulated map as extra input, late-fuse probability features with backbone features, and train with a recursive click-sampling scheme.
Key results
- With ViT-B trained on SBD, MFP achieves NoC85 of 1.39 on Berkeley and 3.92 on DAVIS, versus 1.44 and 4.10 for SimpleClick.
- MFP (ViT-B) reports the best NoC95 on DAVIS (11.27 vs 12.24 for SimpleClick) when trained on SBD.
- Out of 36 settings comparing identical backbones, MFP shows superior results in 28 settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add modulated probability-map input, late fusion, and recursive training (ablation) | ViT-B MFP · DAVIS | NoC@85: 4.10 -> 3.92 | Helped |
| add modulated probability-map input, late fusion, and recursive training (ablation) | ViT-B MFP · DAVIS | NoC@90: 5.60 -> 5.32 | Helped |
| add modulated probability-map input, late fusion, and recursive training (ablation) | ViT-B MFP · DAVIS | NoC@95: 11.58 -> 11.27 | Helped |
| add recursive training instead of the prior training procedure (ablation) | ViT-B MFP · DAVIS | NoC@85: 3.98 -> 3.92 | Helped |
| add recursive training instead of the prior training procedure (ablation) | ViT-B MFP · DAVIS | NoC@90: 5.35 -> 5.32 | No clear effect |
| add recursive training instead of the prior training procedure (ablation) | ViT-B MFP · DAVIS | NoC@95: 11.79 -> 11.27 | Helped |
Limitations
- Gains over the strongest baselines are often small (margins under 0.1 NoC in several settings).
- Relies on fixed hyperparameters (N=7, Rmax=100) chosen without a reported sensitivity study.
Experimental setup
Datasets: SBD COCO LVIS GrabCut Berkeley DAVIS
Benchmarks: GrabCut Berkeley DAVIS SBD
Metrics: NoC85 NoC90 NoC95 AUC of mIoU
Baselines: SimpleClick FocalClick EMC-Click CDNet RITM iCMFormer GPCIS PseudoClick FocusCut BRS f-BRS DynaMITe
Builds on: SimpleClick RITM f-BRS Xception
Closest papers by meaning
- TETRIS: Towards Exploring the Robustness of Interactive Segmentation (AAAI 2024 · similarity 0.86)
- RClicks: Realistic Click Simulation for Benchmarking Interactive Segmentation (NeurIPS 2024 · similarity 0.86)
- NTClick: Achieving Precise Interactive Segmentation With Noise-tolerant Clicks (CVPR 2025 · similarity 0.85)
- Repurposing Stable Diffusion Attention for Training-Free Unsupervised Interactive Segmentation (CVPR 2025 · similarity 0.85)
- Order-aware Interactive Segmentation (ICLR 2025 · similarity 0.85)
- Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts (CVPR 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.