AI Paper Atlas

Research map / Label-efficient and robust learning / Semi-supervised learning

Unbiased Subclass Regularization for Semi-Supervised Semantic Segmentation

Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu

CVPR 2022 · Label-efficient and robust learning › Semi-supervised learning · approach: Consistency training for semi-supervised segmentation · 57 citations (OpenAlex)

Check 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 changedWhereMeasuredEffect
Add class-unbiased subclass regularization (ablation)Original-class self-training on USRN · PASCAL VOC 1/32mIoU: 64.1 -> 65.0Improved
Add subclass self-training (ablation)USRN with MSL, OST, and USR · PASCAL VOC 1/32mIoU: +2.1% in mIoUImproved
Add the entropy-based gate mechanism (ablation)USRN with subclass self-training · PASCAL VOC 1/32mIoU: +1.5% in mIoUImproved
Replace balanced k-means with normal k-means (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6 -> 68.0Hurt
Replace balanced k-means with normal k-means (ablation)Subclass annotations · PASCAL VOC 1/32Subclass CBR: 99.5% -> 96.4%Hurt
Remove feature sharing between F and F_sub (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6 -> 67.3Hurt
Share both low- and high-level features (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6 -> 67.8Hurt
Set pseudo-label confidence threshold γ to 0.85 or 0.95 (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6; 68.5No clear effect
Set pseudo-label confidence threshold γ below 0.75 (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6 -> 67.7; 67.4Hurt
Set pseudo-label confidence threshold γ above 0.95 (ablation)USRN · PASCAL VOC 1/32mIoU: 68.6 -> 68.1Hurt
Use USRN instead of the strongest reported baselineSemi-supervised segmentation · PASCAL VOC 1/64; Cityscapes 1/32mIoU: 3.8%; 2.1%Improved
Integrate USRN into existing semi-supervised methodsDARS, CPS, and CAC · PASCAL VOC 1/32mIoU: +4.5; +4.4; +4.9Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy