Research map / Label-efficient and robust learning / Knowledge distillation
Online Knowledge Distillation for Efficient Pose Estimation
Zheng Li, Jingwen Ye, Mingli Song, Ying Huang, Zhigeng Pan
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 |
|---|---|---|---|
| train with OKDHP online distillation | 2-Stack HG · MPII val | PCKh@0.5: 88.6 -> 89.2 | Improved |
| train with OKDHP online distillation | 4-Stack HG · MPII val | PCKh@0.5: 89.2 -> 90.0 | Improved |
| train with OKDHP online distillation | 8-Stack HG · MPII val | PCKh@0.5: 90.2 -> 90.6 | Improved |
| train with OKDHP online distillation | 4-Stack HG · COCO val2017 | AP: 73.6 -> 74.8 | Improved |
| use pixel-wise KL divergence instead of MSE as distillation loss (ablation) | OKDHP 4-Stack HG · MPII val | PCKh@0.5: 89.5 -> 90.0 | Improved |
| use FAU instead of simple mean aggregation of branch heatmaps (ablation) | OKDHP 4-Stack HG · MPII val | PCKh@0.5: 89.7 -> 90.0 | Improved |
| use FAU instead of ONE-style gate module (ablation) | OKDHP 4-Stack HG · MPII val | PCKh@0.5: 89.8 -> 90.0 | Improved |
| increase number of branches from 2 to 5 (ablation) | OKDHP 2-Stack HG · MPII val | PCKh@0.5: 89.2 -> 89.4 | Improved |
| increase individual HG modules per branch from 1 to 4 (ablation) | OKDHP 4-Stack HG · MPII val | PCKh@0.5: 89.8 -> 90.2 | Improved |
| vary loss weight beta from 0.5 to 5 (ablation) | OKDHP 2-Stack HG · MPII val | PCKh@0.5: 89.24 -> 89.18 | No clear effect |
| use unbalanced branches (6- and 8-stack auxiliaries) instead of balanced (ablation) | OKDHP 4-Stack HG · MPII val | PCKh@0.5: 90.0 -> 90.2 | Improved |
| use OKDHP instead of offline FPD distillation | 4-Stack HG · MPII val | PCKh@0.5 / training cost: 89.7 -> 90.0, 66 -> 47 | Improved |
Closest papers by meaning
- Knowledge Distillation for 6D Pose Estimation by Aligning Distributions of Local Predictions (CVPR 2023 · similarity 0.91)
- Fast Human Pose Estimation (CVPR 2019 · similarity 0.89)
- DistilPose: Tokenized Pose Regression With Heatmap Distillation (CVPR 2023 · similarity 0.89)
- Human Pose Estimation via Convolutional Part Heatmap Regression (ECCV 2016 · similarity 0.88)
- Dynamic Kernel Distillation for Efficient Pose Estimation in Videos (ICCV 2019 · similarity 0.87)
- SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation (CVPR 2024 · 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.
