TotalCapture¶
- Modality: Optical motion capture with synchronized multi-view RGB, depth, IMU
- Primary Tasks: 3D pose estimation, multi-view reconstruction, IMU-to-mocap fusion
- Scale: 5 subjects, 4 camera views, 4 Vicon cameras, 11 sequences per subject
- License: Creative Commons Attribution 4.0 (non-commercial recommended)
- Access: http://totalcapture.net/
Summary¶
TotalCapture provides high-quality motion capture data aligned with RGB videos, depth maps, and inertial measurements. It is widely used to evaluate multi-view reconstruction and to benchmark IMU fusion algorithms thanks to precise calibration between sensors and mocap ground truth.
Reference Paper¶
- Matthew Trumble et al. "Total Capture: A 3D Deformation Model for Tracking Faces, Hands, and Bodies." IEEE TPAMI, 2019.
PDF
Benchmarks & Baselines¶
- TotalCapture baseline - MPJPE: 19.0 mm for multi-view reconstruction; Trumble et al., 2019.
- VNect + Fusion - MPJPE: 28.2 mm when fusing IMU data; Mehta et al., SIGGRAPH 2017.
- Standard evaluation splits include Freestyle, Walking, Acting, and Rom sequences, often with cross-subject validation.
Tooling & Ecosystem¶
- TotalCapture Toolbox for downloading and preprocessing.
- OpenPose calibration scripts for aligning 2D detections to 3D.
- IMU-to-mocap fusion frameworks leverage TotalCapture for evaluation.
Known Challenges¶
- Dataset size is modest; combine with AMASS or Human3.6M for training when possible.
- Calibration metadata requires careful parsing; check camera intrinsics/extrinsics before reprojection.
- IMU data contains drift over long sequences; use provided synchronization markers.