Research map / 3D vision / Human pose estimation
Loose Inertial Poser: Motion Capture with IMU-attached Loose-Wear Jacket
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.
Loose Inertial Poser uses four IMUs in a loose-wear jacket and a SeMo-AE autoencoder that models secondary motion as Gaussian noise in latent space to synthesize loose-wear IMU training data for real-time upper-body pose estimation.
Problem
Sparse-IMU motion capture typically requires tight, strapped sensor placement; loose-wear clothing introduces secondary-motion disturbances to IMU signals, and collecting enough loose-wear IMU training data is impractical.
Approach
Learn an autoencoder latent space in which the secondary-motion effect is additive and Gaussian (enforced by an MMD loss on batch-normalized latent differences), then sample Gaussian noise with temporal interpolation to generate diverse synthetic loose-wear IMU data for training a pose estimator.
Key results
- SeMo-AE-augmented model reaches 19.83 deg mean angular error and 10.58 cm positional error, versus 22.28 deg and 15.74 cm for the best baseline PIP.
- Removing noise-guided latent space learning and temporal coherence scheme raises angular error to 21.46 and 21.71 deg respectively in ablation.
- Replacing SeMo-AE with a conditional GAN gives 25.06 deg angular error, barely better than no synthetic data (24.18 deg).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Loose Inertial Poser instead of PIP | Sparse inertial motion capture on loose-wear clothing · DSreal Loose, all participants | angular error (degree): 22.28 ± 6.28 -> 19.83 ± 6.79 | Helped |
| Use Loose Inertial Poser instead of PIP | Sparse inertial motion capture on loose-wear clothing · DSreal Loose, all participants | positional error (cm): 15.74 ± 7.80 -> 10.58 ± 6.24 | Helped |
| Add noise-guided latent space learning (ablation) | SeMo-AE in Loose Inertial Poser · DSreal Loose, all participants | angular error (degree): 21.54 ± 6.52 -> 19.83 ± 6.79 | Helped |
| Add noise-guided latent space learning (ablation) | SeMo-AE in Loose Inertial Poser · DSreal Loose, all participants | positional error (cm): 11.91 ± 6.33 -> 10.58 ± 6.24 | Helped |
| Add Temporal Coherence Scheme (ablation) | SeMo-AE in Loose Inertial Poser · DSreal Loose, all participants | angular error (degree): 21.46 ± 7.19 -> 19.83 ± 6.79 | Helped |
| Add Temporal Coherence Scheme (ablation) | SeMo-AE in Loose Inertial Poser · DSreal Loose, all participants | positional error (cm): 12.11 ± 7.23 -> 10.58 ± 6.24 | Helped |
| Replace SeMo-AE with cGAN for synthetic IMU generation | Loose Inertial Poser · DSreal Loose, all participants | angular error (degree): 19.83 ± 6.79 -> 25.06 ± 7.18 | Hurt |
| Replace SeMo-AE with cGAN for synthetic IMU generation | Loose Inertial Poser · DSreal Loose, all participants | positional error (cm): 10.58 ± 6.24 -> 16.47 ± 7.24 | Hurt |
| Train directly on simulated loose-wear data without synthetic augmentation | Pose estimation network · DSreal Loose, all participants | angular error (degree): 19.83 ± 6.79 -> 24.18 ± 7.38 | Hurt |
| Train directly on simulated loose-wear data without synthetic augmentation | Pose estimation network · DSreal Loose, all participants | positional error (cm): 10.58 ± 6.24 -> 18.61 ± 6.53 | Hurt |
Surprising findings
- cGAN-generated loose-wear data provided little improvement over training on the simulated loose data alone.
- Baseline SOTA methods retrained on simulated loose data perform worse than their tight-wear designs suggest, with PIP producing abnormal joint position estimates.
Limitations
- Only one garment was produced, so the impact of clothing size and fabric is unexplored.
- Generated data does not cover all wearing styles, causing higher errors for unzipped wear.
- Arm twist rotation is hard to measure accurately with loose-wear IMUs; the jacket is not washable.
Experimental setup
Datasets: AMASS TailorNet simulated clothing data
Benchmarks: Loose Inertial Poser real-world loose-wear test dataset
Metrics: angular error (degrees) positional error (cm)
Baselines: DIP IMUPoser TransPose TIP PIP cGAN
Builds on: TransPose TailorNet MMD conditional GAN
Closest papers by meaning
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.86)
- UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units (CVPR 2025 · similarity 0.85)
- MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis (CVPR 2025 · similarity 0.85)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.84)
- DLCA-Recon: Dynamic Loose Clothing Avatar Reconstruction from Monocular Videos (AAAI 2024 · similarity 0.83)
- HMD-Poser: On-Device Real-time Human Motion Tracking from Scalable Sparse Observations (CVPR 2024 · similarity 0.82)
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.