Research map / Label-efficient and robust learning / Semi-supervised learning
When Confidence Fails: Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation
Pan Liu, Jinshi Liu
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 residual dispersion and optimized selection (M_i) to semi-supervised baseline (ablation) | SemiBaseline · PASCAL VOC original 1/16 | mIoU: 72.9 -> 73.8 | Improved |
| add residual dispersion and optimized selection (M_i) to semi-supervised baseline (ablation) | SemiBaseline · PASCAL VOC original 1/2 | mIoU: 78.6 -> 79.3 | Improved |
| add smooth loss weight on top of M_i (ablation) | CSL · PASCAL VOC original 1/16 | mIoU: 73.8 -> 74.2 | Improved |
| add smooth loss weight on top of M_i (ablation) | CSL · PASCAL VOC original 1/2 | mIoU: 79.3 -> 79.8 | Improved |
| add reliable-pixel masking loss on top of M_i and smooth weights (ablation) | CSL · PASCAL VOC original 1/16 | mIoU: 74.2 -> 76.8 | Improved |
| add reliable-pixel masking loss on top of M_i and smooth weights (ablation) | CSL · PASCAL VOC original 1/2 | mIoU: 79.8 -> 80.9 | Improved |
| compute pseudo-label selection on labeled plus unlabeled data (D_l U D_u) instead of unlabeled only (ablation) | CSL · PASCAL VOC original 1/16 | mIoU: 76.8 -> 75.6 | Hurt |
| use class-specific CSL instead of class-agnostic (ablation) | CSL · PASCAL VOC original 1/16 | mIoU: 76.8 -> 74.6 | Hurt |
| combine all four metrics (v_n, entropy, residual entropy, margin) instead of confidence plus residual dispersion (ablation) | CSL · PASCAL VOC original 1/8 | mIoU: 79.6 -> 71.5 | Hurt |
| add residual entropy as extra metric (ablation) | CSL · PASCAL VOC original 1/8 | mIoU: 79.6 -> 77.3 | Hurt |
| add margin as extra metric (ablation) | CSL · PASCAL VOC original 1/8 | mIoU: 79.6 -> 76.9 | Hurt |
| use loss weights lambda1=1.00, lambda2=1.00 instead of 0.50/0.50 (ablation) | CSL · PASCAL VOC original 1/16 | mIoU: 76.8 -> 72.5 | Hurt |
Closest papers by meaning
- Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels (CVPR 2022 · similarity 0.91)
- Using Unreliable Pseudo-Labels for Label-Efficient Semantic Segmentation (IJCV 2024 · similarity 0.91)
- PseudoSeg: Designing Pseudo Labels for Semantic Segmentation (ICLR 2021 · similarity 0.90)
- Beyond Confidence: Exploiting Homogeneous Pattern for Semi-Supervised Semantic Segmentation (ICML 2025 · similarity 0.88)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.88)
- Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters (NeurIPS 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.
