Begin with the deployment input
A model processing an ongoing stream faces transitions, idle time, and uncertain task boundaries. A model receiving preselected task clips faces a different problem. State which input your project assumes before deciding on collection units. Trimmed footage can be useful, but it should not be described as equivalent to continuous activity merely because all clips came from one session.
Retain source relationships
If clips are derived from longer recordings, keep the original session identifier, source time interval, and transformation history. This allows related examples to be grouped and labels to be traced. Overlapping clips may share most of their visual content even when their names differ. Review those relationships before counting clips as independent examples or dividing them for evaluation.
Define trimming rules
Explain whether clips include preparation, transitions, and the visible result. A fixed window around one action can truncate a long attempt or include a neighboring task. Provide examples for short and long activities, interruptions, and simultaneous actions. Decide how a clip is labeled when its boundaries cut through an action that began earlier in the source.
Review storage and sampling choices
Long sessions contain uneven activity density, while task clips can overrepresent visually obvious actions. Request summaries of active time, task episodes, and unused intervals when those distinctions matter. For Fiund sourcing inquiries, give the desired review unit and target coverage rather than only a footage duration. That makes the relationship between volume and usable context more explicit.
Review checklist
- Describe the model input unit
- Preserve session and interval references
- Document trimming conventions
- Report coverage beyond clip totals