Research map / 3D vision / Human pose estimation
An Empirical Study of the Collapsing Problem in Semi-Supervised 2D Human Pose Estimation
Rongchang Xie, Chunyu Wang, Wenjun Zeng, Yizhou 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 |
|---|---|---|---|
| use easy-hard augmentation pair with gradient only through the hard branch instead of plain consistency (ablation) | consistency-based SSL · COCO val (1K labels) | AP: 0.9 -> 41.5 | Improved |
| detach the teacher (easy) signal so gradients flow only through the student (ablation) | Ours · COCO | collapse | Improved |
| use easy augmentation for both branches or hard for the teacher instead of easy-hard (ablation) | Ours · COCO | heatmap response | Hurt |
| learn dual networks instead of a single network (ablation) | Ours · COCO val (1K labels) | AP: 38.5 -> 41.5 | Improved |
| add Joint Cutout as hard augmentation to dual model (ablation) | Ours (Dual) · COCO val (1K labels) | AP: 41.5 -> 44.6 | Improved |
| add Joint Cutout to single model (ablation) | Ours (Single) · COCO val (1K labels) | AP: 38.5 -> 42.1 | Improved |
| use Joint Cutout instead of RandAug as hard augmentation (ablation) | Ours (Dual) · COCO val (1K labels) | AP: 43.7 -> 44.6 | Improved |
| apply hard augmentation in supervised training only (ablation) | Supervised ResNet18 · COCO val | AP: 51.1 -> 52.1 | Mixed |
| pair ResNet18 with HRNet w48 partner instead of another ResNet18 (ablation) | Ours (Dual) · COCO val (1K labels) | AP: 41.5 -> 48.7 | Improved |
| add unlabeled COCO WILD images with many labels | Ours · COCO val, ResNet50 | AP: 70.9 -> 73.9 | Improved |
| add unlabeled COCO WILD images on strong HRNetW48 | Ours · COCO val, HRNetW48 | AP: 77.2 -> 79.2 | Improved |
| add semi-supervised learning to DARK on test-dev | Ours (+DARK) · COCO test-dev | AP: 76.2 -> 77.2 | Improved |
Closest papers by meaning
- 3D Human Pose Estimation in Video With Temporal Convolutions and Semi-Supervised Training (CVPR 2019 · similarity 0.89)
- Multiview-Consistent Semi-Supervised Learning for 3D Human Pose Estimation (CVPR 2020 · similarity 0.88)
- Weakly-Supervised 3D Human Pose Learning via Multi-View Images in the Wild (CVPR 2020 · similarity 0.87)
- Invariant Teacher and Equivariant Student for Unsupervised 3D Human Pose Estimation (AAAI 2021 · similarity 0.87)
- Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions (ECCV 2024 · similarity 0.87)
- 3D Human Pose Estimation From a Single Image via Distance Matrix Regression (CVPR 2017 · similarity 0.87)
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.
