Research map / Label-efficient and robust learning / Semi-supervised learning
Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective
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 ECOCSeg, which replaces one-hot class encoding in pseudo-label learning with error-correcting output codes (ECOC) plus bit-level pseudo-label denoising, improving UDA and SSL segmentation.
Problem
Pseudo-label learning for semantic segmentation in UDA and SSL generates erroneous pseudo-labels that get amplified by one-hot (argmax) encoding during training.
Approach
Encode each class as an ECOC binary codeword (max-min distance or text-based codebooks), predict K bits instead of N-way softmax, and mine reliable shared bits from the nearest-codeword candidates to form hybrid bit-level pseudo-labels, trained with BCE plus pixel-code distance and contrast losses.
Key results
- On GTAv->Cityscapes, ECOCSeg improves DACS by 2.4 mIoU (52.1 to 54.5), DAFormer by 2.2 (68.3 to 70.5), and MIC by 1.0 (75.9 to 76.9).
- On SYNTHIA->Cityscapes, ECOCSeg improves DACS by 3.8 mIoU, DAFormer by 2.4, and MIC by 1.7 (to 69.0).
- SSL gains range from 1.1% to 3.7% across ST++, FixMatch, and UniMatch on Pascal with DeepLabv3+.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ECOCSeg to DACS | DACS · GTAv→Cityscapes val | mIoU: 52.1 -> 54.5 | Helped |
| Add ECOCSeg to DAFormer | DAFormer · GTAv→Cityscapes val | mIoU: 68.3 -> 70.5 | Helped |
| Add ECOCSeg to MIC | MIC · GTAv→Cityscapes val | mIoU: 75.9 -> 76.9 | Helped |
| Add ECOCSeg to DACS | DACS · SYNTHIA→Cityscapes val | mIoU: 48.3 -> 52.1 | Helped |
| Add ECOCSeg to DAFormer | DAFormer · SYNTHIA→Cityscapes val | mIoU: 60.9 -> 63.3 | Helped |
| Add ECOCSeg to MIC | MIC · SYNTHIA→Cityscapes val | mIoU: 67.3 -> 69.0 | Helped |
| Use BCE alone instead of the full customized optimization criterion (ablation) | ECOCSeg with DAFormer · Cityscapes fully supervised | accuracy | Hurt |
| Add pixel-code distance loss (ablation) | ECOCSeg with DAFormer · Cityscapes fully supervised | accuracy: +1.6% | Helped |
| Add pixel-code contrast loss (ablation) | ECOCSeg with DAFormer · Cityscapes fully supervised | accuracy: +1.5% | Helped |
| Combine BCE, pixel-code distance, and pixel-code contrast losses (ablation) | ECOCSeg with DAFormer · Cityscapes fully supervised | accuracy: +0.5% | Helped |
| Use max-min distance encoding instead of text-based encoding (ablation) | ECOCSeg · Cityscapes fully supervised | accuracy | Hurt |
| Use hybrid bit mining labels instead of code-wise or bit-wise labels alone (ablation) | ECOCSeg · UDA pseudo-label learning | performance | Helped |
Limitations
- Error bars are not reported due to computational cost.
- Theoretical analysis relies on independence and uniform random noise assumptions that do not hold exactly in practice.
- Code is promised upon acceptance rather than released with the submission.
Experimental setup
Datasets: GTAv SYNTHIA Cityscapes PASCAL VOC 2012 COCO
Benchmarks: GTAv->Cityscapes SYNTHIA->Cityscapes Pascal VOC 2012 SSL COCO SSL
Metrics: mIoU
Baselines: ProDA CPSL TransDA ADFormer CDAC DACS DAFormer MIC ST++ FixMatch UniMatch
Builds on: DACS DAFormer MIC ST++ UniMatch
Closest papers by meaning
- PseudoSeg: Designing Pseudo Labels for Semantic Segmentation (ICLR 2021 · similarity 0.90)
- Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels (CVPR 2022 · similarity 0.87)
- Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters (NeurIPS 2024 · similarity 0.87)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.86)
- HPL-ESS: Hybrid Pseudo-Labeling for Unsupervised Event-based Semantic Segmentation (CVPR 2024 · similarity 0.86)
- AUCSeg: AUC-oriented Pixel-level Long-tail Semantic Segmentation (NeurIPS 2024 · similarity 0.85)
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.