MINIONS: Motion capture from INertial and visION Sensors

MINIONS is a large-scale multi-modal human motion capture dataset collected from inertial sensors (IMUs) and vision sensors (RGB videos), designed to support research on consumer-affordable motion capture with a monocular camera and sparse IMUs, as well as broader motion understanding tasks.

Key properties

Multi-modalities: captured with up to eight 2K RGB cameras and full-body IMU suits with 17 nine-axis sensors; an RGB-D scanner is used to obtain textured meshes for each actor.
Scalability: 5.5M+ frames and ~440 minutes of action sequences captured from multiple viewpoints.
Diversity: 146 fine-grained action categories, including single-person and multi-person interactions.
Abundance of annotations: time-synchronized labels per frame, including 2D/3D keypoints, joint rotations, SMPL parameters, fine-grained action labels, and texture.

Data acquisition & annotation pipeline (from the paper)

The dataset construction covers the following stages:

Multimedia hardware setup: synchronized RGB cameras, full-body IMUs, and an RGB-D scanner.
Calibration: camera calibration; IMU alignment to the SMPL coordinate system; camera–IMU alignment.
Textured mesh reconstruction: RGB-D scanning to build textured human models and recover approximate SMPL shape parameters.
Global motion annotations: tracking, 2D keypoint detection, 3D triangulation, and motion recovery by combining visual and inertial results, producing high-quality 2D/3D joints and SMPL parameters.

Access policy (controlled access)

The dataset is not released for indiscriminate public download. Access is registration-based and approval-based for eligible users, and requires submitting information and a signed agreement.

Data License: see LICENSE (full text).
Important: the license grants use solely for non-commercial, not-for-profit research and/or non-commercial educational use (see the license for full terms).

Citation

If you use MINIONS, please cite the corresponding paper:

Title: Motion Capture from Inertial and Vision Sensors

@article{minions2026,
  title   = {Motion Capture from Inertial and Vision Sensors},
  author  = {Chen, Xiaodong and Liu, Wu and Bao, Qian and Liu, Xinchen and Dai, Ruoli and Zhang, Yongdong and Mei, Tao},
  journal = {Arxiv},
  year    = {2026}
}