What is markerless motion capture?
No physical markers, no suit, no infrared cameras built for the purpose: markerless motion capture extracts kinematic data straight from ordinary video, using a model that estimates where the joints are in each frame. For how FreeMoCap specifically fits into this picture, see What is FreeMoCap?; this page is about the wider technique it belongs to.
Older than the deep learning boom
The idea predates modern computer vision by decades. In 1983, David Hogg published one of the first attempts to model a person from video: a walking figure represented as a hierarchy of cylinders, fitted to the image. In 1996, Davis and Gavrila built on that idea with one of the earliest multi-camera markerless systems, mapping an articulated cylindrical model onto images of a real person.
Without a way to directly recognize anatomical features in an image, these early systems had to work from indirect cues instead: the outline of a silhouette, optical flow between frames, edge detection, or reconstructing a 3D visual hull from multiple camera views. Real methods, but ones that demanded serious computer vision expertise to build and tune.
The shift to learned models
That changed with deep learning-based pose estimation. DeepPose, released in 2014, and OpenPose, released in 2017, marked the transition. Instead of hand-built geometric models, a neural network learns to infer joint locations directly from an image. Deep learning quickly became the dominant approach, and it's what every actively developed markerless system, FreeMoCap included, is built on today.
Three tradeoffs, rarely solved together
Modern markerless motion capture research and tooling tends to fall into one of three categories:
- Custom-built workflows, developed by individual research groups for a specific study. Described in a paper's methods section, but the code is usually not shared, so other researchers rebuild it from scratch.
- open source systems meant for reuse, like OpenCap (kinematics from two or more iPhones), Pose2Sim, and PosePipe. These lower the barrier to entry but are often harder to adapt: swapping the underlying pose estimation model or changing the processing pipeline isn't always straightforward.
- Commercial proprietary systems, most notably Theia3D. Accurate and widely validated, but expensive and closed source, which limits transparency into how a number was actually calculated.
Across all three, the same pattern shows up: gains in accuracy, accessibility, or adaptability tend to come at the expense of one of the other two. It's the gap FreeMoCap was built to close, not by finding a clever workaround, but by treating accessibility itself as a design constraint that can improve data quality rather than compromise it. See How FreeMoCap works for how that plays out in practice.