Separate activities from actions
Preparing a meal, opening a drawer, and grasping a utensil describe different levels of behavior. Decide whether your project needs all three and how they relate. A hierarchy can preserve detail without requiring every example to share one level of specificity. Start from the intended model output instead of collecting every verb annotators can imagine.
Write distinctions with examples
Provide short examples for neighboring labels such as moving, placing, and arranging. Include near misses where a hand approaches an object without completing the action. Define whether the noun describes the contacted object, the tool, or the destination. This prevents apparently consistent labels from referring to different parts of the interaction across annotation batches.
Handle uncertainty and overlap
People can carry one item while opening a door or use both hands on separate objects. State whether multiple labels may overlap in time. Keep an unknown or not-visible option when the footage does not support a fine distinction. Forcing a detailed label can create false certainty that is difficult to detect through simple category counts.
Pilot the vocabulary before expansion
Have reviewers label the same varied examples and compare the reasons for disagreement. Revise definitions that depend on hidden intent or inconsistent action boundaries. Preserve a vocabulary version and a mapping for renamed categories. When requesting additional footage, include the current definitions so sourcing and annotation use the same idea of what each target behavior means.
Review checklist
- Define the action hierarchy
- Show neighboring-label examples
- Allow overlap and uncertainty
- Version label definitions and changes