Research map / 3D vision / Human pose estimation
SRNet: Improving Generalization in 3D Human Pose Estimation with a Split-and-Recombine Approach
Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, Stephen Lin
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 low-dimensional global context recombination to the split-fuse-split network (ablation) | SFS · Human3.6M Subject protocol | MPJPE (mm): 39.4 -> 36.6 | Improved |
| add low-dimensional global context recombination to the split-fuse-split network (ablation) | SFS · Human3.6M Rare Pose 20% | MPJPE (mm): 59.2 -> 48.6 | Improved |
| add low-dimensional global context recombination to the split-fuse-split network (ablation) | SFS · Human3.6M Rare Pose 10% | MPJPE (mm): 68.2 -> 53.7 | Improved |
| combine late fusion and early split (SFS) instead of fully connected (ablation) | FC · Human3.6M Rare Pose 10% | MPJPE (mm): 89.1 -> 68.2 | Improved |
| combine late fusion and early split (SFS) instead of fully connected (ablation) | FC · Human3.6M Subject protocol | MPJPE (mm): 46.8 -> 39.4 | Improved |
| add global context back to local groups (H=1) versus none (group-connected only) (ablation) | GP · Human3.6M Subject protocol | MPJPE (mm): 62.7 -> 38.3 | Improved |
| reduce global context dimension from 200% of local feature to a single dimension (ablation) | SRNet · Human3.6M Rare Pose 20% | MPJPE (mm): 78.5 -> 49.1 | Improved |
| replace all fully connected layers with group-connected layers (no cross-group information) (ablation) | LF/ES networks · Human3.6M Rare Pose 20% | MPJPE (mm) | Hurt |
| use more than five local joint groups (ablation) | SRNet · Human3.6M | MPJPE (mm) | Hurt |
| randomly shuffle joints among groups instead of physically local groups (ablation) | SRNet · Human3.6M | MPJPE (mm) | Hurt |
| use additive instead of concatenation combination operator (ablation) | SRNet · Human3.6M Subject protocol (2D GT) | MPJPE (mm): 38.3 -> 36.4 | Improved |
| use multiplication instead of concatenation combination operator (ablation) | SRNet · Human3.6M Subject protocol (2D GT) | MPJPE (mm): 38.3 -> 36.6 | Improved |
Later papers that compared with this method
3 later accepted papers report SRNet as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- FMPose3D: monocular 3D pose estimation via flow matching (CVPR 2026 · better on 1 of 1 shared result)
- Learning Skeletal Graph Neural Networks for Hard 3D Pose Estimation (ICCV 2021 · better on 4 of 4 shared results)
- 3D Human Pose Estimation With Spatial and Temporal Transformers (ICCV 2021 · 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
- Adversarial PoseNet: A Structure-Aware Convolutional Network for Human Pose Estimation (ICCV 2017 · similarity 0.90)
- Human Pose Estimation Using Global and Local Normalization (ICCV 2017 · similarity 0.90)
- A Simple yet Effective Baseline for 3D Human Pose Estimation (ICCV 2017 · similarity 0.90)
- 3D Human Pose Estimation From a Single Image via Distance Matrix Regression (CVPR 2017 · similarity 0.90)
- Locally Connected Network for Monocular 3D Human Pose Estimation (TPAMI 2020 · similarity 0.89)
- A Dual-Augmentor Framework for Domain Generalization in 3D Human Pose Estimation (CVPR 2024 · similarity 0.88)
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.
