Research map / Label-efficient and robust learning / Semi-supervised learning
MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis
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 Motion-Drift Augmentation (MODA), an IMU-specific augmentation that simulates body-worn sensor drift (self-rotation and sliding) and evaluates it on sparse-IMU human pose estimation and IMU action recognition across five settings.
Problem
Data augmentation for IMU signals is underexplored, and generic time-series augmentations (jittering, moving average, masking, flipping, slicing) ignore IMU properties such as rotation constraints and motion consistency, hurting human motion analysis.
Approach
Model IMU displacement during motion as time-varying rotational and translational offsets between sensor and bone frames, and synthesize new IMU readings by applying randomized drift to the bone-to-sensor transforms, so augmented data stays on SO(3) and remains physically consistent with the motion.
Key results
- On CIP full-supervised, MODA reduces Angular error from 10.01 to 8.02 and SIP error from 14.96 to 13.29 versus the baseline.
- On BABEL-150 full-supervised, MODA improves top-1 accuracy from 44.87% to 48.71% (+3.84%) and top-5 by 8.45 points.
- In semi-supervised CIP with FixMatch, MODA with 1/8 labels reaches Angular error 8.44 versus 10.01 for the full-label baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MODA | IMU-based HPE · CIP, fully supervised | Angular error: 10.01 -> 8.02 | Helped |
| Add MODA | IMU-based HPE · ANDY, fully supervised | Angular error: 5.20 -> 4.81 | Helped |
| Add MODA | IMU-based HPE · CIP, domain adaptation | Positional error: 6.99 -> 6.36 | Helped |
| Add MODA | IMU-based HPE · ANDY, domain adaptation | Positional error: 6.15 -> 4.65 | Helped |
| Add MODA | IMU-based HPE · DIP, domain adaptation | Positional error: 8.58 -> 7.67 | Helped |
| Add MODA | IMU-based HPE · TotalCapture, domain adaptation | Positional error: 9.40 -> 6.17 | Helped |
| Add MODA | IMU-based HPE · CIP, domain generalization | Positional error: 9.78 -> 9.17 | Helped |
| Add MODA | IMU-based HPE · ANDY, domain generalization | Positional error: 9.64 -> 7.72 | Helped |
| Add MODA | IMU-based HPE · DIP, domain generalization | Positional error: 10.30 -> 8.25 | Helped |
| Add MODA | IMU-based HPE · TotalCapture, domain generalization | Positional error: 8.74 -> 8.23 | Helped |
| Add MODA | IMU-based HAR · BABEL-150, fully supervised | top-1 accuracy: 44.87% -> 48.71% | Helped |
| Add MODA | IMU-based HAR · BABEL*-150, few-shot | top-1 accuracy: 40.09% -> 41.23% | Helped |
Surprising findings
- Generic augmentations such as moving average and jittering often degrade IMU HPE and HAR performance, sometimes below the no-augmentation baseline.
- Under domain generalization, models trained only on synthetic data produce much jitterier motion on real data (jitter rising about 50% over domain adaptation).
Limitations
- Evaluated only on sparse 6-IMU HPE and on a synthetic HAR dataset (BABEL); domain adaptation and generalization were not tested for HAR due to lack of public datasets.
- Few-shot setting was not evaluated for HPE because of limited motion data per class.
Experimental setup
Datasets: AMASS DIP TotalCapture CIP ANDY BABEL
Benchmarks: CIP ANDY DIP TotalCapture BABEL
Metrics: SIP error Angular error Positional error Mesh error Jitter top-1 accuracy top-5 accuracy
Baselines: Jittering Moving Average Random Mask Flipping Slicing PoseAugment Motion Mirroring FixMatch
Builds on: FixMatch PoseAugment Transpose
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)
- Loose Inertial Poser: Motion Capture with IMU-attached Loose-Wear Jacket (CVPR 2024 · similarity 0.85)
- Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation (NeurIPS 2024 · similarity 0.84)
- Spatial-Related Sensors Matters: 3D Human Motion Reconstruction Assisted with Textual Semantics (AAAI 2024 · similarity 0.84)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.84)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.84)
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.