Research map / Label-efficient and robust learning / Semi-supervised learning
Learning from Future: A Novel Self-Training Framework for Semantic Segmentation
Read the paperCheck 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 |
|---|---|---|---|
| use future (virtually updated) student for teacher update, deep K=3 (FST-D) | self-training DeepLabV2 R101 · GTAV -> Cityscapes | mIoU: 56.3 -> 59.8 | Helped |
| naive FST: use only the virtual future student to update teacher (ablation) | ST · GTAV -> Cityscapes | mIoU: 56.3 -> 56.4 | No clear effect |
| keep current student in teacher update (improved FST, K=1) (ablation) | naive FST · GTAV -> Cityscapes | mIoU: 56.4 -> 57.7 | Helped |
| explore wider (FST-W) with the same data batch (ablation) | ST · GTAV -> Cityscapes | mIoU: 56.3 -> 56.8 | No clear effect |
| use different data batches for FST-W exploration (ablation) | FST-W · GTAV -> Cityscapes | mIoU: 56.8 -> 59.3 | Helped |
| increase serial exploration steps K from 2 to 3 (ablation) | FST-D · GTAV -> Cityscapes | mIoU: 57.8 -> 59.8 | Helped |
| increase K from 3 to 4 (ablation) | FST-D · GTAV -> Cityscapes | mIoU: 59.8 -> 59.7 | No clear effect |
| vary future momentum mu' from 0.99 to 0.99999 (ablation) | FST · GTAV -> Cityscapes | mIoU: 58.8 -> 59.9 | No clear effect |
| train ST 4x longer instead of using FST (ablation) | ST · GTAV -> Cityscapes | mIoU: 56.3 -> 59.3 | Helped |
| increase ST batch size to 4x (ablation) | ST · GTAV -> Cityscapes | mIoU: 56.3 -> 55.5 | Hurt |
| train ST 3x longer on DAFormer (ablation) | ST DAFormer · GTAV -> Cityscapes | mIoU: 68.3 -> 68.3 | No clear effect |
| add FST on semi-supervised PASCAL VOC with PSPNet at 1/16 labels | ST · PASCAL VOC 1/16 | mIoU: 65.47 -> 68.35 | Helped |
Closest papers by meaning
- ST++: Make Self-Training Work Better for Semi-Supervised Semantic Segmentation (CVPR 2022 · similarity 0.92)
- Semi-Supervised Semantic Segmentation via Gentle Teaching Assistant (NeurIPS 2022 · similarity 0.92)
- Three Ways To Improve Semantic Segmentation With Self-Supervised Depth Estimation (CVPR 2021 · similarity 0.91)
- Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels (CVPR 2022 · similarity 0.91)
- Single-Stage Semantic Segmentation From Image Labels (CVPR 2020 · similarity 0.90)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 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.