AI-generated documentationchecked against source

This page was drafted by an AI assistant from source material and may contain inaccuracies. If you spot something wrong, please open an issue or use the Edit this page link below to submit a fix.

2026-08-26: spot-checked the most load-bearing claims directly: Point3d (skellyforge/data_models/trajectory_3d.py), RealtimeSkeletonRigidifier's docstring (freemocap/core/tasks/mocap/rigid_body/skeleton_rigidifier.py, confirms the enforce_rigid_bones streaming-counterpart claim near-verbatim), CanonicalBodyModelInfo/CanonicalHandModelInfo (skellymodels/models/tracking_model_info.py), interpolate_trajectory/filter_trajectory (post_processing/), and MocapTaskConfig.filter_config: FilterConfig (mocap_task_config.py); all confirmed accurate
2026-08-24: SkellyForge source read directly (full package tree, the pyproject config, README, tracker_info YAMLs); integration points verified in the FreeMoCap clone

How FreeMoCap uses SkellyForge

freemocap/pyproject.toml depends on SkellyForge from GitHub main. The split between direct calls and reimplemented mirrors matters:

Direct calls, the posthoc pipeline. FreeMoCap's mocap helpers do their own triangulation, then hand the result to SkellyForge: wrap the points in a Trajectory3d, run interpolate_trajectory then filter_trajectory, build a Human from MediapipeModelInfo() or RTMPoseModelInfo(), call put_skeleton_on_ground() when the calibration is not ground-plane aligned and fix_hands_to_wrist(), run calculate(), and write all five save-out formats. The same shape produces the Board model from triangulated ChArUco corners. MocapTaskConfig takes its filter settings as SkellyForge's own FilterConfig.

Canonical models and mirrored math, the realtime pipeline. For latency reasons the realtime aggregator does not call SkellyForge's batch functions; instead it loads SkellyForge's canonical models once (CanonicalBodyModelInfo and CanonicalHandModelInfo through AnatomicalStructure.from_model_info, giving the joint hierarchies, bone_length_ratios seeds, and Winter COM table), maps raw detector keypoints onto canonical names using SkellyTracker's per-tracker mapping YAMLs, and reimplements the same algorithms per frame. Its RealtimeSkeletonRigidifier docstring states this plainly: the streaming counterpart of SkellyForge's posthoc enforce_rigid_bones, same median-length plus forward-pass method applied online. Likewise the realtime center_of_mass module follows "the same Winter-table math as SkellyForge" adapted to single-frame inputs.

Shared plumbing. Point3d from skellyforge.data_models.trajectory_3d is the wire type for skeleton points across FreeMoCap's pubsub topics, frontend payloads, and streaming kinematics. The websocket server derives the canonical_body and canonical_hand schemas it sends to the UI from SkellyForge's AnatomicalStructure (landmarks plus segment connections), so the frontend renders 3D skeleton connections without hardcoding them. The startup banner reports the installed SkellyForge version alongside the other Skelly packages.