Recognizing a broad activity and identifying its individual steps require different evidence. Describe the behavior your model should recognize, the context it can access, and the situations in which a prediction matters. Fiund can review a sourcing brief against those requirements; specific collections and permissions need separate assessment.
Choose a useful activity unit
Start with a concrete sequence such as unpacking supplies, preparing a workspace, or putting tools away. Decide whether the label belongs to the whole session, a task episode, or a short action. Include idle periods and transitions if the model must decide when an activity starts rather than receiving pretrimmed examples.
Describe variation that matters
Separate differences in participant, room, objects, and task order. A collection can contain many minutes while repeating one arrangement. Ask for coverage summaries at the level your evaluation needs, and identify conditions that should be reserved for testing before clips are created from longer recordings.
Specify temporal supervision
Action boundaries, activity descriptions, and step sequences answer different questions. Supply examples of ambiguous transitions and overlapping actions. Decide how incomplete tasks should be marked and whether annotators may use later footage when interpreting an earlier moment. That choice matters when the intended model must operate online.
Review representative sequences
Use complete examples to inspect visibility, motion, transitions, and annotation consistency. Review ordinary and difficult footage, not only clean demonstrations. Record which problems are exclusions and which are intentional conditions. Include your intended use so technical fit and the proposed permission scope can be reviewed together. Record whether activities are spontaneous, instructed, or repeated demonstrations, since those collection choices affect how the behavior should be interpreted.
Include in your inquiry
- Activity hierarchy and example sequences
- Context window and prediction timing
- Participant and environment variation
- Annotation examples and evaluation split