AI Paper Atlas

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

Enhanced Soft Label for Semi-Supervised Semantic Segmentation

Jie Ma, Chuan Wang, Yang Liu, Liang Lin, Guanbin Li

ICCV 2023 · Label-efficient and robust learning › Semi-supervised learning · approach: Consistency training for semi-supervised segmentation · 76 citations (OpenAlex)

This paper tried a common change: Add a contrastive loss

add pixel-to-part contrastive loss on top of DSL · Pascal VOC classic Full (1464) · mIoU: 80.41 -> 81.77 Improved

What 1,317 accepted papers report for this change

See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.

Read the paper

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 hard pseudo labels for unlabeled images (ablation)supervised baseline · Pascal VOC classic Full (1464)mIoU: 72.50 -> 78.32Improved
replace hard label with dynamic soft label (ablation)ESL · Pascal VOC classic Full (1464)mIoU: 78.32 -> 80.41Improved
add pixel-to-part contrastive loss on top of DSL (ablation)ESL · Pascal VOC classic Full (1464)mIoU: 80.41 -> 81.77Improved
add pixel-to-part contrastive loss to hard label alone (ablation)hard-label self-training · Pascal VOC classic Full (1464)mIoU: 78.32 -> 80.26Improved
use small eta (0.80) so DSL degrades to hard label (ablation)DSL · Pascal VOC classic Full (1464)mIoU: 80.41 -> 78.43Hurt
use fixed top-5 dominant classes instead of dynamic accumulated-probability threshold (ablation)DSL · Pascal VOC classic Full (1464)mIoU: 80.41 -> 79.89Hurt
raise eta from 0.95 to 0.99 (ablation)DSL · Pascal VOC classic Full (1464)mIoU: 80.41 -> 80.28No clear effect
use K=3 subclass prototypes instead of K=1 (pixel-to-region) (ablation)pixel-to-part contrast · Pascal VOC classic Full (1464)mIoU: 80.74 -> 81.39Improved
increase subclass number K beyond 5 (ablation)pixel-to-part contrast · Pascal VOC classic Full (1464)mIoU: 81.77 -> 81.72 (K=7); 81.64 (K=10)No clear effect
dynamic soft label vs hard label with fewer labels (ablation)ESL · Pascal VOC classic 1/16 (92)mIoU: +5.53 (92 labels) vs +2.09 (1464 labels)Improved
use more labeled data of mixed quality (blender 5291 vs classic 1464)ESL · Pascal VOCmIoU: 81.77 -> 79.98Hurt

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