Research map / Detection and segmentation / Weakly supervised detection and segmentation
W2P: Switching from Weak Supervision to Partial Supervision for Semantic 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 W2P, which reframes the second stage of weakly-supervised semantic segmentation as partially supervised learning with noisy labels, using boundary-preserving noise detection, a teacher-student partially-supervised module, boundary correction and copy-paste boundary generation.
Problem
Pseudo-labels from CAM-based weakly-supervised semantic segmentation are noisy, especially at object boundaries, and training the segmentation network on them limits performance despite improvements in pseudo-label quality.
Approach
Splits pseudo-labels into clean and noisy regions via superpixel-smoothed per-pixel loss and class-wise GMM thresholds, then uses an EMA teacher to relabel noisy regions with confident predictions, with boundary-preserving correction and copy-paste boundary generation to retain boundary supervision.
Key results
- W2P reaches 74.0 mIoU on PASCAL VOC 2012 val (73.9 test) with ImageNet-pretrained ResNet101, outperforming IRN-based methods by 10.7 points over IRN.
- W2P achieves 46.4 mIoU on MS COCO 2014 val, the reported state of the art among compared WSSS methods.
- Boundary mIoU on VOC val rises from 24.7 (IRN) and 33.4 (BECO) to 36.0 with W2P.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply W2P to IRN-generated pseudo-labels (ablation) | W2P with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 65.1 -> 74.0 | Helped |
| Add BPND (ablation) | W2P with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 65.1 -> 67.5 | Helped |
| Add PSL (ablation) | W2P with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 65.1 -> 70.7 | Helped |
| Add M+CRF refinement to BPND and PSL (ablation) | W2P with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 73.1 -> 74.0 | Helped |
| Add boundary-preserving noise correction (BPNC) (ablation) | PSL with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 70.5 -> 71.9 | Helped |
| Add adaptive thresholding to BPNC (ablation) | PSL with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 71.9 -> 73.6 | Helped |
| Add boundary generation (ablation) | PSL with IRN pseudo-labels · PASCAL VOC 2012 val | mIoU: 73.6 -> 74.0 | Helped |
| Use adaptive thresholding instead of fixed-percentage selection for BPND (ablation) | W2P · PASCAL VOC 2012 val | mIoU: 61.0, 70.9, 69.5 -> 74.0 | Helped |
| Use adaptive thresholding instead of fixed-percentage selection for PSL (ablation) | W2P · PASCAL VOC 2012 val | mIoU: 72.8, 73.6, 73.0 -> 74.0 | Helped |
| Set τ1 to 0.9 rather than the tested alternatives (ablation) | W2P BPND · PASCAL VOC 2012 val | mIoU: 72.2, 73.8, 74.0, 73.2, 67.6 | Mixed |
| Set τ2 to 0.9 rather than the tested alternatives (ablation) | W2P PSL · PASCAL VOC 2012 val | mIoU: 73.4, 73.1, 74.0, 73.6, 73.0 | Mixed |
| Set EMA coefficient λ to 0.99 rather than the tested alternatives (ablation) | W2P · PASCAL VOC 2012 val | mIoU: 70.3, 71.0, 73.7, 74.0, 73.9 | Mixed |
Surprising findings
- Methods that improve pseudo-label mIoU (e.g., AMN over EDAM) do not necessarily improve the final segmentation val mIoU, showing a mismatch between label quality and segmentation performance.
Limitations
- Relies on superpixels (SLIC) to preserve boundaries, which depends on low-level feature clustering quality.
- Uses multiple hyperparameters (tau1, tau2, EMA lambda) that need tuning, though reported as minimal.
Experimental setup
Datasets: PASCAL VOC 2012 MS COCO 2014
Benchmarks: PASCAL VOC 2012 MS COCO 2014
Metrics: mIoU boundary mIoU mAcc
Baselines: IRN ReCAM AMN ADELE BECO URN AEFT ViT-PCM ToCo MCTformer SANCE DRS
Builds on: IRN ReCAM AMN ADELE BECO Mean Teacher
Closest papers by meaning
- Boundary-Enhanced Co-Training for Weakly Supervised Semantic Segmentation (CVPR 2023 · similarity 0.93)
- Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach (AAAI 2020 · similarity 0.92)
- Threshold Matters in WSSS: Manipulating the Activation for the Robust and Accurate Segmentation Model Against Thresholds (CVPR 2022 · similarity 0.91)
- Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.90)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Separate and Conquer: Decoupling Co-occurrence via Decomposition and Representation for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
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.