Publish each episode's fitted source-world ground plane for baseline, synthetic random-walk, and CARI4D reconstruction variants. Declare the sidecar in episode metadata and document the coordinate convention so SOMA-to-G1 retargeting preserves the released training reference.
Increase the deterministic synthetic random-walk bounds to 16 mm / 8 degrees for the human root and 30 mm / 30 degrees for the object. The resulting 3.815 cm CD-h and 6.045 cm CD-o are calibrated to half of CARI4D's BEHAVE paper errors, with the exact 8,196-point evaluation embedded in provenance.
Conjugate the first-frame human-mesh transform from MHR-native coordinates into the OpenCV world used by object poses, and retain the baseline canonical object mesh frame. The provenance includes the coordinate basis and CARI4D-style CD-h/CD-o metrics.
Apply one first-frame human-mesh-derived rigid transform to every human and object frame, and express object poses in the baseline canonical mesh frame. The provenance records all episode transforms, object-frame transforms, hashes, and residuals.
Reorganize the existing Track 2 dataset under baseline/ and add complete synthetic random-walk and CARI4D reconstruction variants with shared synchronized videos.
