A first-person recording can show a hand reaching toward an object without showing the eventual contact or result. For interaction research, define which moments must be observable and what each annotation should describe. The sourcing conversation should connect camera placement to the specific evidence your model needs.
Specify the interaction of interest
Distinguish reaching, touching, grasping, moving, using, and releasing an object. Identify whether both hands and multiple objects may participate. A hand near a tool is not sufficient evidence of tool use; provide positive examples and near misses that explain the difference for reviewers and annotators.
Capture the relevant object states
If the task concerns change, request context before and after the interaction. Opening a container and moving a closed container need different labels. Explain whether state is determined visually, through a task record, or through another signal, and allow an uncertain outcome when the evidence is obscured.
Define spatial annotations carefully
Boxes, masks, tracks, and contact relations require different review effort and provide different supervision. Name the active object and explain how to handle an object that is temporarily hidden. Require stable references across frames if the intended task depends on following the same instance over time.
Evaluate visibility across the task
Inspect reaching below the camera, work near the torso, and interactions at the edge of the image. Decide whether natural occlusion belongs in the dataset or whether critical events must remain visible. Human footage can support interaction understanding, but it does not inherently contain executable robot commands. For small objects, inspect the delivered image rather than relying on recording resolution alone; cropping and compression can change usable detail.
Include in your inquiry
- Interaction vocabulary and object categories
- Required before-and-after context
- Visibility and occlusion policy
- Spatial labels and instance identity rules