Research map / 3D vision / Human pose estimation
Human Body Model Fitting by Learned Gradient Descent
Jie Song, Xu Chen, Otmar Hilliges
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 current parameter estimate Theta as network input (ablation) | learned update network (x + gradient) · Human3.6M (gradient-input ablation) | Reconstruction error (mm): 64.2 -> 56.4 | Improved |
| add gradient as network input (ablation) | learned update network (x + Theta) · Human3.6M (gradient-input ablation) | Reconstruction error (mm): 61.9 -> 56.4 | Improved |
| add target 2D pose x as input (ablation) | learned update network (Theta + gradient) · Human3.6M (gradient-input ablation) | Reconstruction error (mm): 60.5 -> 56.4 | Improved |
| add Theta and gradient inputs to x-only network (ablation) | residual lifting network · Human3.6M (gradient-input ablation) | Reconstruction error (mm): 67.7 -> 56.4 | Improved |
| increase update iterations from 1 to 4 (ablation) | ours · Human3.6M (Protocol 2) | Reconstruction error (mm): 66.3 -> 56.6 | Improved |
| increase update iterations from 4 to 5 (ablation) | ours · Human3.6M (Protocol 2) | Reconstruction error (mm): 56.6 -> 56.4 | No clear effect |
| add gradient-based iterative refinement versus direct lifting (ablation) | direct 2D-to-3D lifting · Human3.6M (Protocol 2) | Reconstruction error (mm): 67.8 -> 56.4 | Improved |
| replace SMPLify optimization with learned gradient descent | SMPLify · 3DPW | Reconstruction error (mm): 106.1 -> 55.9 | Improved |
| replace SMPLify optimization with learned gradient descent | SMPLify · Human3.6M (Protocol 2) | Reconstruction error (mm): 82.3 -> 56.4 | Improved |
| replace SMPLify-X with learned gradient descent | SMPLify-X · EHF | Reconstruction error (mm): 57.6 -> 54.7 | Improved |
| replace regression training (SPIN with additional 3D data) with learned fitting | SPIN · Human3.6M (Protocol 2) | Reconstruction error (mm): 41.1 -> 56.4 | Hurt |
Closest papers by meaning
- Learned Vertex Descent: A New Direction for 3D Human Model Fitting (ECCV 2022 · similarity 0.93)
- Learning to Reconstruct 3D Human Pose and Shape via Model-Fitting in the Loop (ICCV 2019 · similarity 0.92)
- Neural Descent for Visual 3D Human Pose and Shape (CVPR 2021 · similarity 0.91)
- Convolutional Mesh Regression for Single-Image Human Shape Reconstruction (CVPR 2019 · similarity 0.90)
- Learning to Estimate 3D Human Pose and Shape From a Single Color Image (CVPR 2018 · similarity 0.90)
- Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (NeurIPS 2024 · similarity 0.89)
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.
