Research map / 3D vision / Human pose estimation
Modeling Human Motion with Quaternion-Based Neural Networks
Dario Pavllo, Christoph Feichtenhofer, Michael Auli, David Grangier
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 |
|---|---|---|---|
| model velocities instead of absolute rotations (ablation) | QuaterNet (RNN) · Human3.6M short-term, proposed protocol, average, 80 ms | mean angle error: 0.45 -> 0.37 | Improved |
| model velocities instead of absolute rotations (ablation) | QuaterNet (RNN) · Human3.6M short-term, proposed protocol, average, 400 ms | mean angle error: 1.17 -> 1.14 | Improved |
| add scheduled sampling (vs teacher forcing) (ablation) | QuaterNet vel. (RNN) · Human3.6M short-term, proposed protocol, average, 400 ms | mean angle error: 1.23 -> 1.14 | Improved |
| add scheduled sampling (vs teacher forcing) (ablation) | QuaterNet vel. (RNN) · Human3.6M short-term, proposed protocol, average, 80 ms | mean angle error: 0.35 -> 0.37 | Hurt |
| add scheduled sampling (vs teacher forcing) (ablation) | QuaterNet CNN vel. · Human3.6M short-term, proposed protocol, average, 400 ms | mean angle error: 1.30 -> 1.19 | No clear effect |
| use recurrent instead of convolutional architecture | QuaterNet vel. · Human3.6M short-term, proposed protocol, average, 400 ms | mean angle error: 1.19 -> 1.14 | Improved |
| use recurrent instead of convolutional architecture | QuaterNet abs. · Human3.6M short-term, proposed protocol, average, 400 ms | mean angle error: 1.55 -> 1.17 | Improved |
| evaluate with 128 samples per sequence instead of 4 (ablation) | zero-velocity baseline · Human3.6M average error variance (25th-75th quantile) | mean angle error: 0.395 -- 0.435 -> 0.413 -- 0.420 | Improved |
| model absolute rotations instead of velocities for long horizon (ablation) | QuaterNet · Human3.6M long-term (up to 1 s) | mean angle error | Improved |
| use forward-kinematics position loss instead of angle loss (ablation) | QuaterNet · Locomotion generation | position error | Improved |
| use quaternions instead of Euler angles or axis-angle (ablation) | QuaterNet · Locomotion generation (long-term) | position and velocity error | Improved |
| increase conditioning length beyond 10-20 frames (ablation) | QuaterNet · Human3.6M Walking, 80 ms | angle error | No clear effect |
Closest papers by meaning
- QuaMo: Quaternion Motions for Vision-based 3D Human Kinematics Capture (ICLR 2026 · similarity 0.90)
- Algebraically Rigorous Quaternion Framework for the Neural Network Pose Estimation Problem (ICCV 2023 · similarity 0.88)
- Structured Prediction Helps 3D Human Motion Modelling (ICCV 2019 · similarity 0.88)
- Towards Natural and Accurate Future Motion Prediction of Humans and Animals (CVPR 2019 · similarity 0.87)
- Asymptotically Stable Quaternion-valued Hopfield-structured Neural Network with Periodic Projection-based Supervised Learning Rules (NeurIPS 2025 · similarity 0.87)
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · 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.
