Name the purpose of each view
A participant camera can reveal their local working perspective, while an external camera may show body movement or parts of the task outside that view. Decide what each stream contributes to the model or reviewer. More cameras do not automatically create more useful supervision if they omit the required evidence or record different activities.
Distinguish paired from unrelated footage
Two recordings of the same task category are not necessarily paired observations of the same event. If correspondence matters, request shared episode identifiers and a documented timing relationship. Define how pauses, missing segments, and restarts are represented. Review an actual matched sequence so the meaning of paired footage is established by evidence rather than a directory structure.
Specify additional geometry only if needed
Some tasks require information about camera placement or calibration; others need only contextual views for human review. State the requirement explicitly and ask what measurements are available. Do not infer geometric alignment from visual similarity. If cameras or mounts move during capture, explain how that change affects any supplied spatial information and the labels derived from it.
Plan inputs and evaluation boundaries
Decide whether the model receives every view or only the egocentric stream. An annotator may use an external view to resolve an ambiguous event, but that evidence should be identified if unavailable to the model. Group all views and derived clips from the same episode when creating evaluation partitions so one event does not leak across the intended boundary.
Review checklist
- Assign a purpose to every view
- Verify episode and timing correspondence
- State calibration needs explicitly
- Group related views for evaluation