Research map / Federated learning and privacy / Differential privacy
Differentially Private 2D Human Pose Estimation
Kaushik Bhargav Sivangi, Paul Henderson, Fani Deligianni
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 subspace projection to DP-SGD (ablation) | DP-SGD finetuning · MPII (C=0.01, eps=0.2) | mean PCKh@0.5: 63.85 -> 78.48 | Improved |
| add subspace projection to DP-SGD (ablation) | DP-SGD finetuning · MPII (C=0.01, eps=0.8) | mean PCKh@0.5: 78.17 -> 80.63 | Improved |
| add subspace projection to DP-SGD (ablation) | DP-SGD training from scratch · MPII (C=0.1) | mean PCKh@0.5: 9.80 -> 13.05 | Improved |
| add subspace projection to DP-SGD (ablation) | DP-SGD finetuning from scratch · MPII (C=0.1) | mean PCKh@0.5 | No clear effect |
| add feature-level privacy (FDP) to DP-SGD (ablation) | DP-SGD finetuning · MPII (C=0.01, eps=0.2) | mean PCKh@0.5: 63.85 -> 75.46 | Improved |
| add feature-level privacy (FDP) to DP-SGD (ablation) | DP-SGD finetuning · MPII (C=0.01, eps=0.8) | mean PCKh@0.5: 78.17 -> 80.40 | Improved |
| combine feature-level privacy and subspace projection (vs DP-SGD) (ablation) | DP-SGD finetuning · MPII (C=1.0, eps=0.8) | mean PCKh@0.5: 12.53 -> 71.66 | Improved |
| combine feature-level privacy and subspace projection (vs FDP alone) (ablation) | FDP training from scratch · MPII (C=0.1, eps=0.8) | mean PCKh@0.5: 11.22 -> 33.48 | Improved |
| combine feature-level privacy and subspace projection (vs DP-SGD) (ablation) | DP-SGD training from scratch · MPII (C=0.1, eps=0.8) | mean PCKh@0.5: 6.85 -> 33.48 | Improved |
| lower clipping threshold from C=0.1 to C=0.01 (ablation) | DP-SGD finetuning · MPII (eps=0.2) | mean PCKh@0.5: 28.46 -> 63.85 | Improved |
| lower clipping threshold from C=1.0 to C=0.01 (ablation) | DP-SGD finetuning · MPII (eps=0.2) | mean PCKh@0.5: 5.94 -> 63.85 | Improved |
| apply DP updates to a frozen backbone instead of finetuning from scratch | Feature-Projective DP · HumanART | AP: 69.5 -> 63.3 non-private; reverses under DP | Mixed |
Closest papers by meaning
- Don’t call it privacy-preserving or human-centric pose estimation if you don’t measure privacy (NeurIPS 2025 · similarity 0.90)
- Learning Privacy-Preserving Optics for Human Pose Estimation (ICCV 2021 · similarity 0.88)
- Private-kNN: Practical Differential Privacy for Computer Vision (CVPR 2020 · similarity 0.86)
- Privacy Preserving Image Queries for Camera Localization (ICCV 2019 · similarity 0.86)
- Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation (ICLR 2025 · similarity 0.84)
- ProbPose: A Probabilistic Approach to 2D Human Pose Estimation (CVPR 2025 · similarity 0.84)
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.
