ViPoser: Sparse-IMU Based Human Pose Estimation with Distilled Vision Foundation Priors
Published in ACM International Conference on Mobile Computing and Networking (MobiCom 2026), 2026
Estimating full-body pose from only one to three consumer-grade IMUs is highly under-constrained, and prior methods often produce physically implausible configurations. ViPoser analyzes the human-centric vision model Sapiens to locate the layers richest in biomechanical and skeletal knowledge (blocks 18–21), distills them into a compact 64-dim student Transformer, and integrates the distilled priors into a three-stage pipeline (local joint estimation → masked whole-body completion → global translation).
Highlights
- 10–30% lower pose error than IMUPoser and MobilePoser across DIP-IMU, TotalCapture, and the IMUPoser dataset (e.g., 13.94 cm MPJVE on DIP-IMU).
- Introduces the Joint Violation Rate (JVR) metric based on AAOS clinical range-of-motion limits; ViPoser reaches near-zero JVR (0.03%), an order of magnitude below baselines.
- Runs at ~9,200 FPS on a consumer RTX 3070 — ~7.7× faster than MobilePoser — suitable for real-time edge deployment.
Recommended citation: Hanyu Zeng, et al. (2026). "ViPoser: Sparse-IMU Based Human Pose Estimation with Distilled Vision Foundation Priors." ACM MobiCom 2026.
Download Paper