Skip to main content
Version: 1.x (legacy)

Choose a tracking model

FreeMoCap can run on several different pose estimation backends without changing anything else in your pipeline (see SkellyTracker for how that swapping works). This page is about which one to actually reach for.

If You're Not Sure, Start With MediaPipe​

MediaPipe is the default for a reason: it's free, runs on a CPU with no GPU required, and is simple to install and call from Python. Unless you have a specific reason to switch, it's the right starting point.

Match the Backend to Your Task​

The full comparative results are in Accuracy, validity, and limits; the short version, by task:

  • Gait. RTMPose was the most consistent backend in FreeMoCap's validation study, with the lowest ankle-kinematics error and none of MediaPipe's speed-dependent lag.
  • Balance, or anything close to static. MediaPipe's built-in temporal smoothing, a liability for fast movement, is what made it the only backend sensitive enough to detect the expected changes in postural sway. RTMPose and ViTPose showed more frame-to-frame jitter at rest.
  • Fast or ballistic movement. Be aware of MediaPipe's smoothing-driven lag and ViTPose's small systematic scaling bias (2 to 4 percent, concentrated in the vertical axis) before relying on either for precise timing.
  • A population or body type standard models weren't trained on. General-purpose backends can fail outright, as they did tracking a transfemoral prosthesis in FreeMoCap's own validation work. A custom-trained model may be the only option that works; see below.

When the Built-In Options Aren't Enough​

If none of the standard backends track your subject reliably, SkellyTracker's standardized interface means a custom-trained model (DeepLabCut, for example) can be integrated without rewriting the rest of the pipeline. That's exactly what FreeMoCap's prosthetics validation case required: off-the-shelf models, trained mostly on able-bodied people, didn't generalize to a prosthetic limb, and a custom-trained model did. See Image tracking and pose models for how that integration works.

Next steps​