docs/recipes/hand-object-3d-hoi.md

Hand-Object 3D HOI

Goal

Model hands, objects, contact, and 3D pose from first-person video — for AR/VR manipulation, dexterous robot learning, or fine-grained interaction understanding.

Start with

• HOT3D (Aria + Quest 3, 3D hand/object/camera poses).
• HOI4D (2.4M RGB-D frames, category-level 4D HOI).
• H2O (two hands + objects, 6D pose, meshes).
• ARCTIC (bimanual, articulated objects, contact).

Add if you need

• Hand-object segmentation: EgoHOS, VISOR.
• First-person hand action / pose: FPHA, Ego2HandsPose.
• In-the-wild capture: SHOW3D.
• Contact / pressure: EgoTouch / TouchAnything.
• Emerging sensors: EgoEMG, EgoEVHands (event cameras).

Benchmarks

HOT3D Challenge (3D hand-object tracking), HOI4D tasks, ARCTIC, and EgoHOIBench / EgoNCE++ for open-vocabulary HOI. Choose by what you predict: poses/meshes (HOT3D/ARCTIC), category-level 4D (HOI4D), or contact (ARCTIC, EgoTouch).

Baselines / tools

• HOT3D tooling and HOI4D tooling for loading poses/meshes/point clouds.
• EgoHOS model for hand-object segmentation.
• Project Aria Tools for Aria streams/calibration.

Minimum viable experiment

1. Pick HOT3D (poses) or HOI4D (segmentation/4D) and its official metric.
2. Reproduce a provided baseline on the standard split.
3. Add your component (e.g., a temporal smoother or contact head).
4. Report the official metric plus a contact/occlusion breakdown.

Common pitfalls

• Synthetic vs. real domain gaps (UnrealEgo, EgoGTA) — report which you trained on.
• Occlusion and motion blur dominate egocentric HOI; bin metrics by occlusion.
• 6DoF object pose needs correct camera intrinsics/extrinsics — verify calibration.
• Mesh/contact licenses (e.g., ARCTIC is request) — confirm before redistribution.

Reporting checklist

• Dataset, split, and calibration source.
• Metric matches the official protocol (e.g., MPJPE, contact F1, AUC).
• Real vs. synthetic training data stated.
• Occlusion/blur breakdown reported.
• License/access of any redistributed annotations.