Research map / Label-efficient and robust learning / Semi-supervised learning
Improving Semantic Segmentation via Efficient Self-Training
Yi Zhu, Zhongyue Zhang, Chongruo Wu, Zhi Zhang, Tong He, Hang Zhang, Raghaven Manmatha, Mu Li, Alexander Johannes Smola
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 pseudo labels from Cityscapes coarse images | Self-training student · Cityscapes val | mIoU: 78.1 -> 79.0 | Improved |
| add pseudo labels from Mapillary images as well | Self-training student · Cityscapes val | mIoU: 78.1 -> 79.3 | Improved |
| add centroid sampling (ablation) | Self-training student · Cityscapes val | mIoU: 78.9 -> 80.0 | Improved |
| increase pseudo-to-real ratio from 1:1 to 7:1 (ablation) | Self-training student with centroid sampling · Cityscapes val | mIoU: 79.3 -> 80.0 | Improved |
| increase pseudo ratio to 7:1 without centroid sampling (ablation) | Self-training student without centroid sampling · Cityscapes val | mIoU: 78.9 -> 78.7 | Hurt |
| replace real-label duplication with pseudo labels (ablation) | Self-training student · Cityscapes val | mIoU: 79.1 -> 80.0 | Improved |
| use self-training instead of Mapillary labeled pre-training | DeepLabV3+ WideResNet38 · Cityscapes val | mIoU: 81.5 -> 82.2 | Improved |
| add warm-up with crop 800 for first 20 epochs to coarse2fine+ schedule (ablation) | Fast training schedule · Cityscapes val | mIoU: 79.5 -> 80.0 | Improved |
| use soft pseudo labels instead of hard (ablation) | Self-training student · Cityscapes val | mIoU: 80.0 -> 79.6 | Hurt |
| pre-train on pseudo labels then finetune instead of joint learning (ablation) | Self-training student · Cityscapes val | mIoU: 80.0 -> 79.2 | Hurt |
| add extra teacher-student loops (2-loop, 3-loop) (ablation) | Self-training student · Cityscapes val | mIoU: 80.0 -> 79.9 -> 80.2 | No clear effect |
| use a better teacher (WideResNet38, 80.5 vs 78.1) (ablation) | Student models · Cityscapes val | mIoU: +0.2 to +0.5 | Improved |
Closest papers by meaning
- Learning from Future: A Novel Self-Training Framework for Semantic Segmentation (NeurIPS 2022 · similarity 0.94)
- Three Ways To Improve Semantic Segmentation With Self-Supervised Depth Estimation (CVPR 2021 · similarity 0.93)
- ST++: Make Self-Training Work Better for Semi-Supervised Semantic Segmentation (CVPR 2022 · similarity 0.92)
- Rethinking Self-Supervised Semantic Segmentation: Achieving End-to-End Segmentation (TPAMI 2024 · similarity 0.92)
- Improving Semi-Supervised and Domain-Adaptive Semantic Segmentation with Self-Supervised Depth Estimation (IJCV 2023 · similarity 0.92)
- Re-Distributing Biased Pseudo Labels for Semi-Supervised Semantic Segmentation: A Baseline Investigation (ICCV 2021 · similarity 0.91)
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.
