Wear and Go: Calibration Free Sparse Inertial Motion Capture via Online Sensor Setup Estimation

Yuanshuo Tan1·Licheng Zhong2

1 Shanghai Jiao Tong University2 National University of Singapore

arXiv

Overview

Abstract

Sparse inertial motion capture typically assumes that each sensor's anatomical identity and mounting orientation are known. We present MOSAIC (Motion-based Sensor Assignment and Inertial Calibration), a framework that recovers this setup from ordinary motion without a calibration pose. Our central insight is that body pose changes continuously while the sensor setup persists, allowing movements that are individually ambiguous to provide complementary evidence over time. A setup network processes unordered sensor streams with temporal and cross-sensor attention, then accumulates assignment and rotation evidence into a session-level estimate. The recovered setup canonicalizes measurements for frozen pretrained pose estimators, separating sensor configuration from motion reconstruction. For sensors worn at prescribed body sites, evaluation on real inertial recordings with randomized setups shows near-calibrated pose accuracy for bounded offsets and substantial benefits from accumulation under arbitrary orientations. Transfer across pose networks supports setup recovery as a reusable interface between interchangeable sensors and pretrained pose models.

Method

MOSAIC recovers sensor identities and mounting orientations from ordinary motion using six IMUs at prescribed body sites. It accumulates evidence across motion windows and corrects the measurements for existing pose networks, without requiring a calibration pose.

MOSAIC: encode unordered sensor streams, accumulate assignment and rotation evidence, then canonicalize measurements for pose reconstruction.

Public dataset results

Real IMU recordings with simulated arbitrary sensor configurations. Root-aligned poses at 1× speed.

Live demos

Our Xsens recordings with arbitrary sensor setups, at 1× speed. Original IMU pose predictions are aligned to the visual reference for display.

Qualitative comparison

Supplementary comparison on our collected recordings: RGB, visual reference, no calibration, TIC, invariant poser, calibrated input, and MOSAIC.
Our collected IMU data, with GVHMR as the visual reference. Calibrated input uses known sensor identities and the initial visual pose and IMU readings, with heading aligned over the first 10 s.

Quantitative results

Per-take
Setup estimation starts afresh for each recording.
Per-session
Setup evidence carries across a subject's recordings under one fixed sensor configuration.

θmax is the largest allowed mounting-angle offset from the calibrated sensor orientation. Arbitrary mounting allows any 3D orientation.

Angular error (°) ↓

DIP-IMU

    TotalCapture

      Pose accuracy over time

      Angular-error gaps to calibrated input over 240 seconds on DIP-IMU and TotalCapture, comparing MOSAIC with bounded and arbitrary mountings against the invariant poser.
      Angular-error gap to calibrated input over the first four minutes, under per-session evaluation. Zero means the same ground-truth-referenced error as calibrated input.

      Mounting recovery over time

      Stacked twist and swing errors for the forearms, shanks, head, and pelvis on DIP-IMU and TotalCapture at 5 seconds, 10 seconds, and averaged after 10 seconds and one minute.
      Per-segment twist and swing errors under arbitrary mountings, per session. Bars show 5 s, 10 s, and averages after 10 s and 1 min. Stacked heights are not total rotation errors.

      BibTeX

      @article{tan2026wearandgo,
        title   = {Wear and Go: Calibration Free Sparse Inertial Motion Capture via Online Sensor Setup Estimation},
        author  = {Tan, Yuanshuo and Zhong, Licheng},
        journal = {arXiv preprint arXiv:2611.22894},
        year    = {2026}
      }