Research map / 3D vision / Human pose estimation
Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DynaIP, a two-stage part-based sparse-IMU human pose estimator that regresses pseudo-velocities and trains on unified real inertial mocap datasets across skeleton formats instead of synthetic AMASS data.
Problem
Full-body human pose estimation from six sparse IMUs suffers from motion ambiguity, underuse of acceleration, and poor generalization because training relies on synthetic virtual IMU data from AMASS with a gap to real measurements.
Approach
Maps real inertial mocap datasets with different skeleton formats (Xsens to SMPL) into one training set via one-to-one bone orientation mapping, learns pseudo-velocities of IMU joints in a first stage, and splits the body into three regions (upper limbs, torso, lower limbs) with a global feature extractor for coherence in a second pose-regression stage.
Key results
- On DIP-IMU, DynaIP* reduces global pose error by 19% relative to PIP (Ang Err 7.07 vs 8.73 deg per Table 2 comparison, as stated in text).
- On Natural Motion, DynaIP reduces SIP error by 28% relative to PIP when trained only on Xsens data (15.78 vs 19.62 deg).
- Training on real Xsens data alone reduces SIP error by 22% on the DIP-IMU full set versus AMASS-only training (18.98 vs 24.50 deg).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DynaIP instead of PIP | IMU-based pose estimation trained on real inertial mocap data · DIP-IMU, AnDy, UNIPD, CIP, Natural Motion | SIP error (°): 17.62 -> 17.31; 9.49 -> 8.93; 8.90 -> 7.29; 12.68 -> 11.42; 19.62 -> 15.78 | Helped |
| Use DynaIP instead of PIP | IMU-based pose estimation trained on real inertial mocap data · DIP-IMU, AnDy, UNIPD, CIP, Natural Motion | Angular error (°): 8.33 -> 7.66; 4.09 -> 3.45; 3.59 -> 2.77; 5.52 -> 4.54; 7.49 -> 7.18 | Helped |
| Add part-based modeling (ablation) | DynaIP* · DIP-IMU, CIP, Natural Motion | SIP error (°): 14.97 -> 13.67; 13.00 -> 11.67; 31.62 -> 18.88 | Helped |
| Add part-based modeling (ablation) | DynaIP* · DIP-IMU, CIP, Natural Motion | Angular error (°): 6.02 -> 5.83; 4.86 -> 4.63; 9.60 -> 8.03 | Helped |
| Add pseudo-velocity learning (ablation) | DynaIP* · DIP-IMU, CIP, Natural Motion | SIP error (°): 14.87 -> 13.67; 12.54 -> 11.67; 29.15 -> 18.88 | Helped |
| Add pseudo-velocity learning (ablation) | DynaIP* · DIP-IMU, CIP, Natural Motion | Angular error (°): 6.11 -> 5.83; 4.89 -> 4.63; 10.49 -> 8.03 | Helped |
Surprising findings
- Real Xsens-only training (DynaIP) performs comparably to TransPose, which is fine-tuned on DIP-IMU, on the DIP-IMU test set.
Limitations
- Evaluated on sparse IMU datasets with a fixed six-sensor setup
- Relies on a skeleton mapping that drops redundant torso joints, verified only qualitatively via end-effector velocity
Experimental setup
Datasets: DIP-IMU AnDy Emokine Virginia Natural Motion UNIPD CIP AMASS
Benchmarks: DIP-IMU AnDy UNIPD CIP Natural Motion
Metrics: SIP error global angular error position error mesh error
Baselines: DIP TransPose TIP PIP AMASS-trained model
Builds on: PIP TransPose DIP
Closest papers by meaning
- HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units (AAAI 2025 · similarity 0.88)
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.87)
- UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units (CVPR 2025 · similarity 0.87)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.86)
- Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation (NeurIPS 2024 · similarity 0.86)
- MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis (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.