Research map / 3D vision / Human pose estimation
HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units
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 HiPoser, a Mamba-based hierarchical shared-learning model that estimates IMU-based 3D human pose by sequentially predicting torso, lower-limb, upper-limb poses and global translation.
Problem
3D human pose estimation from a sparse set of IMUs suffers from ambiguous local body parts and negative transfer between parts when body-part poses are estimated in parallel from a shared whole-body representation.
Approach
Decomposes IMU-based pose estimation into sequential tasks (torso, lower limbs, upper limbs, global translation) in a hierarchical shared Mamba backbone, feeding each task's features to the next and injecting prior pose/position states at each window; the task order can be changed to prioritize detail or stability.
Key results
- On DIP-IMU (trained on AMASS), HiPoser reaches 14.11 SIP, 8.16 Ang, 5.44 Pos, 6.30 Mesh and 32.48 cm Dist, beating DynaIP (14.97, 8.57, 5.69, 6.50, 33.95).
- On TotalCapture, HiPoser achieves Dist error 36.61 cm versus 40.63 for DynaIP and Pos error 5.63 cm versus 5.96.
- Stability-priority task sequence (Seq-S) gives lower Dist error (29.75 cm vs 32.48 cm) on DIP-IMU, while detail-priority Seq-D gives better pose errors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HiPoser instead of DynaIP | IMU-based 3D HPE · DIP-IMU | Pos error (cm): 5.69 -> 5.44 | Helped |
| use HiPoser instead of DynaIP | IMU-based 3D HPE · TotalCapture | Pos error (cm): 5.96 -> 5.63 | Helped |
| add hierarchical sharing rather than parallel partition modeling (ablation) | HiPoser · DIP-IMU | Dist error (cm): 37.81 -> 32.48 | Helped |
| add hierarchical sharing rather than parallel partition modeling (ablation) | HiPoser · DIP-IMU | SIP error (°): 14.23 -> 14.11 | Helped |
| add body-part partitioning (ablation) | HiPoser · DIP-IMU | Pos error (cm): 5.81 -> 5.44 | Helped |
| add body-part partitioning (ablation) | HiPoser · DIP-IMU | Dist error (cm): 35.49 -> 32.48 | Helped |
| replace whole-body Mamba estimation with hierarchical part-level estimation (ablation) | HiPoser · DIP-IMU | Dist error (cm): 42.48 -> 32.48 | Helped |
| use the stability-priority sequence Upper→Lower→Torso→Trans instead of the detail-priority sequence Torso→Lower→Upper→Trans (ablation) | HiPoser · DIP-IMU | Dist error (cm): 32.48 -> 29.75 | Helped |
| use the detail-priority sequence Torso→Lower→Upper→Trans instead of the stability-priority sequence (ablation) | HiPoser · DIP-IMU | SIP error (°): 14.42 -> 14.11 | Helped |
| add rotation states for all joints and position state (ablation) | HiPoser · DIP-IMU | SIP error (°): 15.60 -> 14.11 | Helped |
| add rotation states for all joints and position state (ablation) | HiPoser · DIP-IMU | Dist error (cm): 47.61 -> 32.48 | Helped |
| replace Mamba with RNN as the backbone (ablation) | HiPoser · DIP-IMU, 20s | Dist error (cm): 62.43 -> 75.68 | Hurt |
Surprising findings
- The stability-priority task order (Upper→Lower→Torso→Trans) yields better translation error than the detail-priority order, while the detail-priority order is better for limb pose.
Limitations
- Translation estimation in the offline setting relies on foot contact probability and velocity combination.
- Evaluated only on IMU-based datasets with a fixed 60Hz sampling rate and fixed window of 300 frames.
Experimental setup
Datasets: AMASS DIP-IMU TotalCapture
Benchmarks: DIP-IMU TotalCapture
Metrics: SIP error Ang error Pos error Mesh error Dist error
Baselines: DIP TransPose TIP PIP DynaIP RNN Bi-RNN Transformer
Builds on: Mamba DynaIP PIP TransPose
Closest papers by meaning
- FisherPoser: Human Motion Estimation from Sparse Observations with Hierarchical Region-Wise Fisher-Matrix Uncertainty Modeling (CVPR 2026 · similarity 0.89)
- HiPART: Hierarchical Pose AutoRegressive Transformer for Occluded 3D Human Pose Estimation (CVPR 2025 · similarity 0.89)
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.88)
- SVTformer: Spatial-View-Temporal Transformer for Multi-View 3D Human Pose Estimation (AAAI 2025 · similarity 0.88)
- Sequential Joint Dependency Aware Human Pose Estimation with State Space Model (AAAI 2025 · similarity 0.87)
- Optimizing Human Pose Estimation Through Focused Human and Joint Regions (AAAI 2025 · similarity 0.87)
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.