Research map / 3D vision / Human pose estimation
OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields
Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, Yaser Ajmal Sheikh
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 PAF-based parsing instead of one-midpoint representation (ablation) | OpenPose · MPII validation (PCKh-0.5) | mAP: 2.9% higher than one-midpoint and 2.3% higher than two intermediate points | Improved |
| add masks of unlabeled persons to the training loss (ablation) | OpenPose · MPII validation | mAP: further improves the performance by 2.3% | Improved |
| use minimal tree edges (13) instead of fully connected graph (91 edges) with greedy parsing (ablation) | OpenPose · MPII custom validation | mAP: 78.3 -> 77.4 (362 s -> 0.005 s per image) | No clear effect |
| train the network only on tree edges (separate training) (ablation) | OpenPose greedy tree · MPII custom validation | mAP: 77.4 -> 79.1 | Improved |
| add 3-scale search to inference (ablation) | OpenPose · MPII full test set | mAP: 72.5 -> 75.6 | Improved |
| add single-person CPM refinement to bottom-up output (ablation) | OpenPose original · COCO validation | AP: 58.4 -> 61.0 | Improved |
| replace SSD detector with ground-truth boxes in top-down CPM (ablation) | top-down CPM · COCO validation | AP: 52.7 -> 62.7 | Improved |
| predict PAFs first as prior for confidence maps instead of confidence maps first (ablation) | OpenPose 6-stage · COCO validation | AP: 61.0 -> 65.3 | Improved |
| allocate more stages to PAF branch (5 PAF - 1 CM vs 3 PAF - 3 CM) (ablation) | OpenPose 6-stage · COCO validation | AP: 65.0 -> 65.3 | Improved |
| reduce to 3 PAF - 1 CM (4 stages) (ablation) | OpenPose · COCO validation | AP: 65.3 -> 64.6 | Mixed |
| add foot keypoints to training (body+foot vs body-only) (ablation) | OpenPose · COCO validation | AP: 65.2 -> 65.3 | No clear effect |
| use new body+foot architecture on GPU instead of original model | OpenPose · runtime, CUDA | inference time: 73 ms -> 36 ms | Improved |
Later papers that compared with this method
10 later accepted papers report OpenPose as a baseline in their results tables; 8 of them report a better number on most of the results they share with it.
- Free-Viewpoint Video of Outdoor Sports Using a Drone (ECCV 2024 · better on 1 of 1 shared result)
- NeMo: Learning 3D Neural Motion Fields From Multiple Video Instances of the Same Action (CVPR 2023 · better on 4 of 4 shared results)
- PyMAF-X: Towards Well-Aligned Full-Body Model Regression From Monocular Images (TPAMI 2023 · better on 0 of 1 shared result)
- AlphaPose: Whole-Body Regional Multi-Person Pose Estimation and Tracking in Real-Time (TPAMI 2022 · better on 1 of 1 shared result)
- ZoomNAS: Searching for Whole-body Human Pose Estimation in the Wild (TPAMI 2022 · better on 1 of 1 shared result)
- Keypoint Communities (ICCV 2021 · better on 5 of 5 shared results)
- Whole-Body Human Pose Estimation in the Wild (ECCV 2020 · better on 5 of 5 shared results)
- Single-Network Whole-Body Pose Estimation (ICCV 2019 · better on 4 of 5 shared results)
2 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeeperCut — DeeperCut: A Deeper, Stronger, and Faster Multi-person Pose Estimation Model (ECCV 2016) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields (CVPR 2017 · similarity 0.95)
- Efficient Online Multi-Person 2D Pose Tracking With Recurrent Spatio-Temporal Affinity Fields (CVPR 2019 · similarity 0.89)
- DeepCut: Joint Subset Partition and Labeling for Multi Person Pose Estimation (CVPR 2016 · similarity 0.89)
- Detangling People: Individuating Multiple Close People and Their Body Parts via Region Assembly (CVPR 2017 · similarity 0.89)
- DeeperCut: A Deeper, Stronger, and Faster Multi-person Pose Estimation Model (ECCV 2016 · similarity 0.88)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024 · 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.
