Follow-up to #869 (shipped in PR #909).
The body retarget is correct (upright, tracking — the inversion bug was fixed in caa823de: the harvested standing pose must be stored as an absolute local, since the live Bone::setOrientation path replaces rather than accumulates). Remaining quality gap: arm placement precision is approximate — analytic Pose-IK on a monocular (single-camera) lift has no true depth, so arm bones are aimed from 2D-projected landmark directions and gesture-precise poses land only approximately.
Ideas to explore (not blocking v1):
- Temporal depth priors / stronger One-Euro tuning on the arm chain.
- The SAM 3D Body backend (separate follow-up) regresses true 3D per-joint rotations and would largely address this.
- Optional elbow-hint / twist-reference refinement in
PoseIKSolver.
Validate on Linux with a real camera (the path macOS couldn't exercise: camera access + headless GL both blocked live testing here).
Follow-up to #869 (shipped in PR #909).
The body retarget is correct (upright, tracking — the inversion bug was fixed in
caa823de: the harvested standing pose must be stored as an absolute local, since the liveBone::setOrientationpath replaces rather than accumulates). Remaining quality gap: arm placement precision is approximate — analytic Pose-IK on a monocular (single-camera) lift has no true depth, so arm bones are aimed from 2D-projected landmark directions and gesture-precise poses land only approximately.Ideas to explore (not blocking v1):
PoseIKSolver.Validate on Linux with a real camera (the path macOS couldn't exercise: camera access + headless GL both blocked live testing here).