Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly Supervised Semantic Segmentation via Progressive Confidence Region Expansion
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 PCRE, a one-stage ViT-based WSSS framework that progressively expands high-confidence mask regions (CRME) and refines them with EMA-updated dataset-level class prototypes (CPE) to fix CAM over-expansion.
Problem
Over-expansion of CAM from target objects into background regions when using vision transformers under image-level supervision, producing unreliable pseudo-labels for weakly supervised semantic segmentation.
Approach
Grow a mask from the highest-confidence patches in small steps using a Reference Overlap Ratio selection rule, with patch-class joint confidence and a class-prototype similarity map (updated by EMA from early-stage features) as supervision, so the mask cannot spill into similar background regions.
Key results
- Achieves 47.2% mIoU on COCO val, outperforming one-stage DIAL by 2.8% mIoU and multi-stage PCSS by 1.5%.
- Reaches 77.6% mIoU CAM pseudo-label quality on VOC train, surpassing DuPL (76.0%).
- Obtains 75.5% and 75.9% mIoU on VOC val and test sets, 1.0% above DIAL on both.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Patch-level Confidence Allocation (PCA) (ablation) | PCRE baseline · PASCAL VOC 2012 val | mIoU: 70.7 -> 72.1 | Helped |
| Add Progressive Mask Expansion and Combination (PMEC) (ablation) | PCRE with PCA · PASCAL VOC 2012 val | mIoU: 72.1 -> 74.4 | Helped |
| Add the CPE loss (LCPE) (ablation) | PCRE with CRME · PASCAL VOC 2012 val | mIoU: 74.4 -> 75.5 | Helped |
| Add Progressive Class Prototype Update (PCPU) (ablation) | PCRE with CRME and LCPE · PASCAL VOC 2012 val | mIoU: 75.5 -> 76.3 | Helped |
| Use patch-class joint confidence instead of single confidence (ablation) | CRME confidence allocation · PASCAL VOC 2012 val | mIoU: 73.1 -> 76.3 | Helped |
| Use sum loss instead of max loss (ablation) | Patch-class joint confidence allocation · PASCAL VOC 2012 val | mIoU: 74.5 -> 76.3 | Helped |
| Set ROR thresholds to τlow=0.4 and τhigh=0.5 (ablation) | CRME · PASCAL VOC 2012 val | mIoU: 73.0 -> 73.8 | Helped |
| Use PCRE instead of DIAL | ViT-B/16 WSSS · PASCAL VOC 2012 val | mIoU: 74.5 -> 75.5 | Helped |
| Use PCRE instead of DIAL | ViT-B/16 WSSS · PASCAL VOC 2012 test | mIoU: 74.9 -> 75.91 | Helped |
| Use PCRE instead of PCSS | WSSS · MS COCO 2014 val | mIoU: 45.7 -> 47.2 | Helped |
Limitations
- Gains over the strongest one-stage baselines are moderate (about 1-3 mIoU) and rely on several hyperparameters (thresholds, radius schedule, loss weights).
- Evaluation limited to PASCAL VOC 2012 and MS COCO 2014 with ViT-B/16 backbone.
Experimental setup
Datasets: PASCAL VOC 2012 SBD MS COCO 2014
Benchmarks: PASCAL VOC 2012 MS COCO 2014
Metrics: mIoU over-activation rate (OA) performance ratio vs fully supervised
Baselines: ToCo DuPL DIAL PCSS CTI CLIMS CLIP-ES AFA TSCD MCTformer
Builds on: ToCo DuPL MCTformer CAM
Closest papers by meaning
- Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.91)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Class Tokens Infusion for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
- DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.88)
- Separate and Conquer: Decoupling Co-occurrence via Decomposition and Representation for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
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.