Research map / 3D vision / Human pose estimation
Learned Neural Physics Simulation for Articulated 3D Human Pose Reconstruction
Misha Andriluka, Baruch Tabanpour, Daniel Freeman, Cristian Sminchisescu
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 |
|---|---|---|---|
| train on longer sequences (2 to 20 unrolled steps) (ablation) | LARP dynamics network · Capsule chain dynamics | position error: 1.1 meter -> 11 cm | Improved |
| add joint displacement features and displacement loss (ablation) | LARP dynamics network · Capsule chain dynamics | position error: 15 cm -> 11 cm | Improved |
| supervise positions instead of velocities (ablation) | LARP dynamics network · Capsule chain dynamics | position error: 95 cm -> 11 cm | Improved |
| use gradient clipping for training on sequences longer than 10 steps (ablation) | LARP training · Capsule chain dynamics | training stability | Improved |
| use soft activations (ELU or GeLU) instead of ReLU (ablation) | LARP dynamics network · Capsule chain dynamics | dynamics error | Improved |
| add stop-gradient on contact network output feature (ablation) | LARP contact feature model · Two-chain capsule contact environment | variance across runs | Improved |
| feed contact network output as feature rather than as an impulse in the integrator (ablation) | LARP contact network · Two-chain capsule contact environment | position error | Improved |
| replace Bullet with LARP in physics-based pose reconstruction | video-based 3D pose reconstruction · AIST-hard | MPJPE-G (mm): 654 -> 643 | Improved |
| replace Bullet with LARP in pose reconstruction on H3.6M | video-based 3D pose reconstruction · H3.6M | MPJPE (mm): 84 -> 85 | No clear effect |
| replace Bullet with LARP in pose reconstruction on foot skate | video-based 3D pose reconstruction · H3.6M | Foot skate: 4 -> 5.4 | Hurt |
| replace Bullet or MuJoCo with LARP for parallel simulation | humanoid simulation · 4,096 parallel simulations | time per step (ms): about 20 ms -> 1.3ms | Improved |
Closest papers by meaning
- Contact and Human Dynamics from Monocular Video (ECCV 2020 · similarity 0.89)
- Efficient Differentiable Simulation of Articulated Bodies (ICML 2021 · similarity 0.89)
- Differentiable Dynamics for Articulated 3D Human Motion Reconstruction (CVPR 2022 · similarity 0.88)
- REACTO: Reconstructing Articulated Objects from a Single Video (CVPR 2024 · similarity 0.88)
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.87)
- PhysPT: Physics-aware Pretrained Transformer for Estimating Human Dynamics from Monocular Videos (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.
