Data & collection
Sparse motion capture
Sparse motion capture reconstructs an articulated body's pose from a small set of measurements, such as head, wrist, and ankle tracker poses, instead of directly observing every joint. Unmeasured joints are inferred with kinematic constraints or learned motion priors, so the result is a model-based estimate rather than direct sensor ground truth.
Also known as: sparse full-body tracking, sparse body tracking, sparse motion sensing
Updated
Sparse refers to the measurements, not the output
A sparse system may produce a complete skeleton even though only a few body locations are measured. AvatarPoser reconstructs articulated full-body motion from sparse head-and-hand sensing by using a learned temporal model. Other systems add pelvis or foot trackers, inertial sensors, or explicit kinematic constraints.
The output should not be described as if every joint came from hardware. Head, hand, pelvis, or foot poses may be direct observations while elbows, knees, spine joints, and the root trajectory remain estimates. Datasets should preserve a per-signal provenance that identifies measured anchors, reconstructed joints, and any later smoothing or refinement.
Mount calibration and motion inference are coupled
A tracker measures its own pose. Converting that pose to an anatomical anchor requires a sensor-to-body transform. An incorrect mount offset can be absorbed as an apparently plausible joint rotation, especially when only a few anchors constrain the body.
EgoExplorer addresses this ambiguity by fitting participant geometry and fixed tracker attachments from a prescribed calibration sequence before completing the motion. The calibration remains valid only while the body geometry and mounting configuration still match the fitted profile.
A motion prior supplies possibilities, not ground truth
Several whole-body poses can agree with the same sparse anchors. A learned prior favours motions seen in its training distribution, while joint limits, surface constraints, and ground constraints remove some physically implausible solutions. None of these signals independently proves the hidden joints were recovered exactly.
Validation should state the sensor layout, participants, actions, training overlap, reference system, coordinate alignment, and errors at both measured and unmeasured joints. Results against fitted or model-derived references should be distinguished from independent optical motion capture. Natural recordings without a matched reference can demonstrate coverage and plausibility, but not absolute joint accuracy.
What to retain in a dataset
A reusable sparse-capture record should include raw tracker poses and timestamps, validity flags, tracker serial or mount identity, calibration parameters, body model, coordinate frames, reconstruction code or version, learned-prior checkpoint, constraint settings, and diagnostics or confidence values.
Keeping those items allows another team to re-run the inference, detect a moved sensor, compare reconstruction methods, and avoid treating an estimated joint as a measured label.
Sources
Related terms
Data & collection
Motion capture
Motion capture is the measurement and reconstruction of a person’s or object’s movement over time, commonly as joint positions, orientations or a fitted body model. Optical markers, cameras and inertial sensors can supply the measurements. Humanoid robotics uses the resulting motion sequences for analysis, imitation and retargeting to a robot body.
Models & learning
Pose estimation
Pose estimation is the process of inferring the position and orientation of a body, object, camera, hand, or robot relative to a specified coordinate frame. In three-dimensional robotics this is often called 6D or 6-DoF pose estimation because the result has three translational and three rotational degrees of freedom, even when orientation is stored with more than three numbers.
Data & collection
Sensor calibration
Sensor calibration is the estimation and documentation of parameters that map raw sensor readings into physically meaningful values and known spatial relationships. In robotics it can include scale, bias, distortion, intrinsic camera parameters, sensor-to-sensor or sensor-to-robot transforms, and timing offsets. Calibration does not remove all noise or drift.
Data & collection
Data synchronisation
Data synchronisation is the process of placing sensor, state, action, annotation, and outcome records on a common timeline so samples that describe the same physical instant or transition can be matched. It requires trustworthy timestamps or trigger relationships and an explicit rule for handling streams with different rates, delays, dropped samples, and clock offsets.
Models & learning
SMPL-X
SMPL-X is a learned parametric 3D human model that represents a unified articulated body, hands, and expressive face using shape and pose parameters. It provides a consistent mesh and skeleton, but fitted SMPL-X motion is an estimate derived from observations and model assumptions rather than direct measurement of every joint or surface point.