Research map / Label-efficient and robust learning / Semi-supervised learning
ST++: Make Self-Training Work Better for Semi-Supervised Semantic Segmentation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 strong data augmentations to unlabeled images during re-training (ablation) | Self-training (ST) · Pascal VOC, augmented, 1/8 labeled (1323) | mIoU: 71.4 -> 73.3 | Helped |
| apply strong data augmentations to labeled data as well as unlabeled data (ablation) | Self-training (ST) · Pascal VOC, augmented, 1/8 labeled (1323) | mIoU: 73.3 -> 73.0 | Hurt |
| add Cutout to labeled data (ablation) | Self-training with SDA · Pascal VOC | mIoU: 73.3 vs. 73.2 | Hurt |
| prioritize reliable images by prediction stability, then re-label the remaining images (ablation) | Self-training (ST) · Pascal VOC, augmented | mIoU: 71.6 -> 72.6; 73.3 -> 74.4; 75.0 -> 75.4 | Helped |
| replace reliability-based image selection with random splitting in two-stage re-training (ablation) | ST++ · Pascal VOC, augmented | mIoU: 71.6 -> 71.3; 73.3 -> 73.9; 75.0 -> 74.7 | Mixed |
| select 25% rather than 50% of unlabeled images as reliable (ablation) | ST++ · Pascal VOC, augmented, 1/8 labeled (1323) | mIoU: 74.4 -> 74.0 | Hurt |
| select 75% rather than 50% of unlabeled images as reliable (ablation) | ST++ · Pascal VOC, augmented, 1/8 labeled (1323) | mIoU: 74.4 -> 74.5 | Helped |
| select confident pixels rather than whole reliable images for the first re-training phase (ablation) | ST++ · Pascal VOC, augmented | mIoU: 72.6 vs. 71.3; 74.4 vs. 73.5; 75.4 vs. 74.9 | Hurt |
| add a third iterative re-training stage with newly generated pseudo labels (ablation) | ST++ · Pascal VOC, original training set, 92 labeled images | mIoU: 65.2% to 68.3% | Helped |
| add a third iterative re-training stage with newly generated pseudo labels (ablation) | ST++ · Pascal VOC, original training set, 366 labeled images | mIoU: 74.6% to 75.2% | Helped |
| use image-level prediction stability instead of a pixel-confidence threshold for selection (ablation) | Selective re-training · Pascal VOC, augmented | mIoU: 72.6 vs. 71.3; 74.4 vs. 73.5; 75.4 vs. 74.9 | Helped |
Closest papers by meaning
- Learning from Future: A Novel Self-Training Framework for Semantic Segmentation (NeurIPS 2022 · similarity 0.92)
- Semi-Supervised Semantic Segmentation via Gentle Teaching Assistant (NeurIPS 2022 · similarity 0.90)
- RWMS: Reliable Weighted Multi-Phase for Semi-supervised Segmentation (AAAI 2024 · similarity 0.89)
- AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.88)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · 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.