Research map / Architectures and efficiency / Attention and transformers
Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation
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 a Sequence Structure Module (SSM) that multiplies fixed-length token sequences by a structural matrix (explicit, implicit, or hybrid) before a Transformer, with SSM-S for IMU spatial structure and SSM-T for a smoothness temporal prior, improving sparse-IMU pose estimation accuracy and jitter.
Problem
Native Transformers lack inductive bias for fixed-length sequences with clear spatial and temporal structure, leading to jitter and inaccurate poses in sparse-IMU full-body pose estimation.
Approach
A Sequence Structure Module multiplies the input token sequence by a structural matrix S (explicit from prior knowledge or statistics, learned, or hybrid S_E + P) followed by LayerNorm and MLP, injecting fixed-length structure into a spatial-temporal Transformer; SSM-S uses a correlation matrix from AMASS joint rotations and SSM-T uses a linearly decaying temporal smoothness prior.
Key results
- On DIP-IMU, the method reaches 7.90 SIP error and 6.06 angular error versus 13.71 and 8.75 for the best prior method PNP, with jitter 0.07 vs 0.17.
- On TotalCapture, angular error is 6.82 and jitter 0.09, versus DynaIP's 11.06 and 0.22 (claimed 38% angular-error reduction).
- Ablation on TotalCapture shows SSM-S reduces angular error (8.82 to 7.83) and SSM-T reduces jitter (0.48 to 0.09) when added to the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SSM-S to the spatial-temporal Transformer (ablation) | Spatial-temporal Transformer · TotalCapture | Ang Err: 8.82 -> 7.83 | Helped |
| Add SSM-S to the spatial-temporal Transformer (ablation) | Spatial-temporal Transformer · TotalCapture | Jitter: 0.48 -> 0.43 | Helped |
| Add SSM-T to the spatial-temporal Transformer (ablation) | Spatial-temporal Transformer · TotalCapture | Jitter: 0.48 -> 0.09 | Helped |
| Use EIHS for SSM-S and ES for SSM-T rather than IS for both (ablation) | SSM sequence structures · TotalCapture | Ang Err: 7.79 -> 6.82 | Helped |
| Use ES for SSM-T rather than IS (ablation) | SSM-T · TotalCapture | Jitter: 0.54 -> 0.09 | Helped |
| Use EIHS for SSM-S rather than IS, with ES for SSM-T (ablation) | SSM-S · TotalCapture | Ang Err: 7.94 -> 6.82 | Helped |
| Set temporal structure σ to 10 rather than 20 (ablation) | SSM-T · TotalCapture | Ang Err: 7.76 -> 6.82 | Helped |
| Use Eq. (8) temporal structure instead of the alternative definition marked * (ablation) | SSM-T · TotalCapture | Ang Err: 7.86 -> 6.82 | Helped |
| Use spatial encoder followed by temporal encoder (SE-TE) instead of TE-SE (ablation) | Spatial-temporal Transformer · TotalCapture | Ang Err: 8.67 -> 6.82 | Helped |
| Increase the number of IMUs from 6 to 10 (ablation) | Ours · DIP-IMU | Ang Err: 6.06 -> 4.40 | Helped |
| Remove SSM-S (ablation) | Ours with 10 IMUs · DIP-IMU | Ang Err: 4.40 -> 5.69 | Hurt |
| Remove SSM-T (ablation) | Ours with 10 IMUs · DIP-IMU | Jitter: 0.03 -> 0.13 | Hurt |
Surprising findings
- Explicit-implicit hybrid structure works best for the spatial module, but implicit-only structure for the temporal module yields unacceptable jitter.
- Using the identity-matrix-based implicit structure for SSM-T produced high jitter, while the explicit temporal smoothness prior gave the lowest jitter.
Limitations
- No global translation tracking; relies solely on IMUs which drift.
- Explicit structure construction is only illustrated with two examples and needs further exploration.
- No statistical significance error bars reported.
Experimental setup
Datasets: AMASS DIP-IMU AnDy CIP Emokine
Benchmarks: DIP-IMU TotalCapture AnDy CIP
Metrics: SIP error angular error positional error mesh error jitter
Baselines: DIP TransPose TIP PIP DynaIP PNP GCN + Conv1d
Builds on: Transformer TransPose PIP DynaIP
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)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.85)
- MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis (CVPR 2025 · similarity 0.84)
- Hierarchical State Space Models for Continuous Sequence-to-Sequence Modeling (ICML 2024 · similarity 0.84)
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.83)
- HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units (AAAI 2025 · similarity 0.83)
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.