Research map / 3D vision / Human pose estimation
Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions
Yihao Ai, Yifei Qi, Bo Wang, Yu Cheng, Xinchao Wang, Robby T. Tan
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 Extreme Low-Light Augmentation (ELLA) fake low-light images (ablation) | DEKR-W32 pre-training · ExLPose-test LL-A | AP@0.5:0.95: 0.8 -> 15.8 | Improved |
| add Person-specific Degradation Augmentation (PDA) in stage 2 (ablation) | Single-teacher student · ExLPose-test LL-A | AP@0.5:0.95: 16.8 -> 19.0 | Improved |
| add PDA in stage 2 (ablation) | Single-teacher student · ExLPose-test WL (well-lit) | AP@0.5:0.95: 60.5 -> 59.4 | Hurt |
| add complementary keypoint-based teacher to the main teacher (ablation) | Dual-teacher framework with PDA · ExLPose-test LL-A | AP@0.5:0.95: 19.0 -> 21.1 | Improved |
| add complementary teacher (ablation) | Dual-teacher framework · ExLPose-test WL | AP@0.5:0.95: 59.4 -> 61.6 | Improved |
| use complementary teacher to add pseudo labels beyond main teacher (ablation) | Pseudo-labeling on ExLPose training set · ExLPose training set | valid pseudo labels: 5056 -> +2136 at OKS 0.5 (+49.8%) | Improved |
| train with well-lit labels only plus dual-teacher low-light adaptation versus LSBN+LUPI trained with paired low-light labels | Bottom-up DEKR framework · ExLPose-OCN (avg) | AP@0.5:0.95: 35.2 -> 37.6 | Improved |
| apply low-light image enhancement before pose estimation (RFormer, LLFlow) | Base-well DEKR · ExLPose-OCN (avg) | AP@0.5:0.95: 6.3 -> 18.3 (RFormer), 21.4 (LLFlow) | Improved |
| use generic domain adaptation (UDA-HE) instead of low-light specific augmentation | Base-well DEKR · ExLPose-OCN (avg) | AP@0.5:0.95: 6.3 -> 6.9 | No clear effect |
| compare bottom-up ours against top-down LIME trained with low-light labels | Ours versus LIME on ExLPose-test · ExLPose-test | AP@0.5:0.95: LL-A 21.1 vs 21.1; LL-E 5.0 vs 7.6 | Mixed |
Closest papers by meaning
- Human Pose Estimation in Extremely Low-Light Conditions (CVPR 2023 · similarity 0.92)
- UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation (CVPR 2026 · similarity 0.91)
- Prior-guided Source-free Domain Adaptation for Human Pose Estimation (ICCV 2023 · similarity 0.87)
- An Empirical Study of the Collapsing Problem in Semi-Supervised 2D Human Pose Estimation (ICCV 2021 · similarity 0.87)
- Poseur: Direct Human Pose Regression with Transformers (ECCV 2022 · similarity 0.86)
- ProbPose: A Probabilistic Approach to 2D Human Pose Estimation (CVPR 2025 · 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.
