How the pose estimate works

OpenBikeFit uses MediaPipe Tasks Vision 0.10.35 and Pose Landmarker Lite float16 v1. After explicit camera permission, the model estimates landmarks from frames locally. It supports educational comparison of bicycle setup and repeated 2D movement.

Only landmark indices 11–16 (shoulders, elbows, wrists) and 23–32 (hips, knees, ankles, heels and foot index points), timestamps and quality facts are retained for a pose series. Face and hand landmarks are immediately replaced with zero values. Raw video is not retained by default. Reduced pose remains potentially personal data and is covered by the privacy notice.

• One camera reduces three-dimensional movement to a two-dimensional projection.
• Camera angle/distance, calibration, lighting, clothing, occlusion and rider/bike overlap affect estimates.
• The model may perform differently across bodies, clothing, devices and environments.
• Repeatability can reproduce a systematic error and does not establish accuracy.
• Confidence labels describe tracking/measurement quality; they do not establish scientific or clinical validity.

You start the camera, review the preview and can leave at any time. Low or insufficient quality blocks a mechanical direction and asks for another recording. Settings provides complete export and verified local deletion. The model does not make legal or similarly significant decisions.

The exact artifact hashes and source are recorded in the public project governance pack and third-party notices. A model/runtime change is released only with an updated hash, network capture, documented limitations and claims records.