Select a boundary convention
An action can begin at initial reach, first contact, or visible object movement. None of these choices should be left implicit. Explain which event defines the start and what ends the action. Use a concrete sequence such as reaching for a cup, lifting it, and placing it elsewhere to demonstrate how your rules divide the footage.
Preserve the time reference
Specify whether labels use source-video time, derived-clip time, or frame indices. Document timestamp units and how endpoints are interpreted. If footage is trimmed or re-encoded, retain a mapping to the source. Without that mapping, an otherwise reasonable annotation can land on a different moment after files move through a processing workflow.
Allow realistic complexity
A person may pause while holding an object or overlap two actions. Decide whether pauses belong inside the action interval and whether simultaneous labels are permitted. Distinguish task interruption from normal hesitation. Mark transitions that cannot be located confidently rather than forcing a precise boundary unsupported by the image or the annotation instructions. Show how the rules handle one action immediately reversing another.
Review boundary disagreements
Compare annotations around short actions, partial occlusions, and slow transitions. Ask reviewers to explain which visible cue they used, then refine the convention. Category agreement can look strong while start and end positions remain inconsistent. For sourcing discussions, provide boundary examples and intended evaluation behavior so annotation scope can be assessed with the footage.
Review checklist
- Choose visible start and end cues
- Specify time units and endpoints
- Define pauses and overlapping actions
- Review difficult boundary examples