Research map / 3D vision / Human pose estimation
Unsupervised 3D Human Pose Representation with Viewpoint and Pose Disentanglement
Qiang Nie, Ziwei Liu, Yunhui Liu
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 bidirectional SeBiReNet (shared hidden states) instead of recursive-only tree (ablation) | pose denoising encoder · Cambridge-Imperial APE pose denoising (MLP decoder) | MPJPE (mm): 65.76 -> 42.03 | Improved |
| share hidden states instead of concatenating subnet outputs (ablation) | SeBiReNet · Cambridge-Imperial APE pose denoising (MLP decoder) | MPJPE (mm): 75.17 -> 42.03 | Improved |
| run SeBiReNet recurrently one more time (ablation) | SeBiReNet · Cambridge-Imperial APE pose denoising (MLP decoder) | MPJPE (mm): 42.03 -> 41.58 | No clear effect |
| use SeBiReNet in decoder as well as encoder (ablation) | pose denoising DAE · Cambridge-Imperial APE pose denoising (MLP decoder) | MPJPE (mm): 42.03 -> 33.39 | Improved |
| add denoising autoencoder with corrupted skeletons (ablation) | pose representation · N-UCLA | accuracy (%): 51.53 -> 58.66 | Improved |
| add feature decomposition (FD) into view and pose features (ablation) | raw DAE · N-UCLA | accuracy (%): 58.66 -> 60.61 | Improved |
| add orthogonal regularization L_O on view feature (ablation) | raw DAE + FD · N-UCLA | accuracy (%): 60.61 -> 62.55 | Improved |
| add feature consistency loss L_f between rotated views (ablation) | raw DAE + FD + L_O · N-UCLA | accuracy (%): 62.55 -> 73.81 | Improved |
| add reconstruction loss of randomly rotated pose (ablation) | raw DAE + FD + L_O + L_f · N-UCLA | accuracy (%): 73.81 -> 76.84 | Improved |
| add view-transferred pose losses (full architecture) (ablation) | raw DAE + FD + L_O + L_f + L(R*p) · N-UCLA | accuracy (%): 76.84 -> 80.30 | Improved |
| add feature consistency loss L_f (ablation) | raw DAE + FD + L_O · NTU RGB+D | accuracy (%): 74.46 -> 75.72 | Improved |
| add view-transferred pose losses (full architecture) (ablation) | raw DAE + FD + L_O + L_f + L(R*p) · NTU RGB+D | accuracy (%): 77.07 -> 79.71 | Improved |
Closest papers by meaning
- Unsupervised 3D Pose Estimation With Geometric Self-Supervision (CVPR 2019 · similarity 0.88)
- Kinematic-Structure-Preserved Representation for Unsupervised 3D Human Pose Estimation (AAAI 2020 · similarity 0.88)
- Weakly-Supervised Discovery of Geometry-Aware Representation for 3D Human Pose Estimation (CVPR 2019 · similarity 0.88)
- CoDePose: Multi-View 3D Human Pose Estimation via Coupled 2D-3D Denoising Diffusion (ECCV 2026 · similarity 0.87)
- Self-Supervised 3D Human Pose Estimation via Part Guided Novel Image Synthesis (CVPR 2020 · similarity 0.87)
- Disentangled Diffusion-Based 3D Human Pose Estimation with Hierarchical Spatial and Temporal Denoiser (AAAI 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.
