Research map / Label-efficient and robust learning / Semi-supervised learning
Pixel Contrastive-Consistent Semi-Supervised Semantic Segmentation
Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, Yu-Xiong Wang
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 pixel contrastive loss to label consistency training (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 63.75 -> 64.63 | Improved |
| add label consistency loss to pixel contrastive loss (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 53.53 -> 64.63 | Improved |
| sample negatives from different images and debias with pseudo labels instead of uniformly (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 62.70 -> 64.63 | Improved |
| add Different Image constraint to uniform sampling (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 62.70 -> 63.33 | Improved |
| use ground-truth Oracle negatives instead of Diff+Pseudo sampling (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 64.63 -> 64.78 | No clear effect |
| increase number of negative pixels from 200 to 800 (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 64.63 -> 65.03 | Improved |
| use L2 label consistency loss instead of cross-entropy with pixel contrastive (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 58.47 -> 64.63 | Improved |
| use pixel contrastive instead of image-level contrastive loss (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 62.49 -> 64.63 | Improved |
| apply contrastive loss at Conv5 instead of Conv3 (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 60.70 -> 64.63 | Improved |
| apply contrastive loss at both Conv5 and Encoder instead of Conv5 alone (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 64.63 -> 64.59 | No clear effect |
| use separate projection heads instead of shared (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 63.97 -> 64.63 | Improved |
| start unlabeled losses from step 0 instead of delaying 2,000 steps (ablation) | PC2Seg · VOC 2012 val, 1/8 labeled (ResNet-50) | mIoU: 63.30 -> 64.63 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PseudoSeg — PseudoSeg: Designing Pseudo Labels for Semantic Segmentation (ICLR 2021) · better on 8 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Semi-Supervised Semantic Segmentation With Directional Context-Aware Consistency (CVPR 2021 · similarity 0.94)
- Semi-supervised Semantic Segmentation with Prototype-based Consistency Regularization (NeurIPS 2022 · similarity 0.92)
- Semi-Supervised Semantic Segmentation With Pixel-Level Contrastive Learning From a Class-Wise Memory Bank (ICCV 2021 · similarity 0.92)
- Querying Labeled for Unlabeled: Cross-Image Semantic Consistency Guided Semi-Supervised Semantic Segmentation (TPAMI 2023 · similarity 0.91)
- Semi-Supervised Semantic Segmentation With High- and Low-Level Consistency (TPAMI 2019 · similarity 0.91)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · 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.
