Research map / 3D vision / Human pose estimation
GFPose: Learning 3D Human Pose Prior With Gradient Fields
Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, 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 the score-based generative pose estimator rather than the deterministic counterpart (ablation) | Human3.6M pose estimation · Human3.6M Protocol #1, GT 2D train/test | MPJPE (mm): 41.9 -> 35.6 | Improved |
| Increase hypotheses from S=10 to S=200 | GFPose · Human3.6M Protocol #1 | minMPJPE (mm): 45.1 -> 35.6 | Improved |
| Generate 200 pose hypotheses instead of using the strongest listed baseline | Human3.6M multi-hypothesis estimation · Human3.6M Protocol #1 | minMPJPE (mm): 44.3 -> 35.6 | Improved |
| Add part-level condition masking with p_p=0.1 (ablation) | GFPose · Human3.6M Protocol #1, single hypothesis | minMPJPE (mm): 51.0 -> 50.5 | Improved |
| Add joint-level condition masking with p_j=0.1 (ablation) | GFPose · Human3.6M Protocol #1, single hypothesis | minMPJPE (mm): 51.0 -> 51.7 | Hurt |
| Use mixed condition masking in unified U2D rather than no masking (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 35.6 -> 36.7 | Hurt |
| Train with a unified U3D model that also conditions on 3D poses (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 36.7 -> 35.9 | Mixed |
| Reduce sampling steps from 1k to 0.5k (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 35.6 -> 36.1 | Hurt |
| Increase sampling steps from 1k to 2k (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 35.6 -> 35.7 | No clear effect |
| Increase score-network depth from 2 to 3 residual blocks (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 35.6 -> 37.2 | Hurt |
| Reduce hidden dimension from 1024 to 512 (ablation) | GFPose · Human3.6M Protocol #1, 200 hypotheses | minMPJPE (mm): 35.6 -> 44.3 | Hurt |
| Augment a small real-pose training set with 2,438 generated poses | Deterministic pose estimator · Human3.6M Protocol #1, 2,438 real poses | MPJPE (mm): 66.1 -> 60.0 | Improved |
Later papers that compared with this method
1 later accepted paper reports GFPose as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Occlusion Handling in 3D Human Pose Estimation with Perturbed Positional Encoding (ECCV 2024 · better on 3 of 3 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
- Pose-NDF: Modeling Human Pose Manifolds with Neural Distance Fields (ECCV 2022 · similarity 0.88)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.88)
- RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency (ECCV 2024 · similarity 0.88)
- HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units (AAAI 2025 · similarity 0.87)
- A Dual-Augmentor Framework for Domain Generalization in 3D Human Pose Estimation (CVPR 2024 · similarity 0.86)
- PoseGen: Learning to Generate 3D Human Pose Dataset with NeRF (AAAI 2024 · similarity 0.86)
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.
