General annotation workflow

Before you start

Read the prompt and the relevant type guide, then watch the full video. If multiple views are supplied, review every view before deciding that an action, occlusion, or episode is unusable.

Four steps

1. Understand the task: confirm the type, output format, timestamp format, and coverage required.
2. Watch and record: follow people, objects, actions, contacts, and visible effects in time order.
3. Enter the answer: answer only what the prompt asks and preserve its required structure.
4. Review and submit: check evidence, completeness, format, and required fields; save, then submit.

Always avoid

• Inferring content, intent, speech, or hidden objects from common sense.
• Importing facts from another camera view when the task requires a single view.
• Editing the prompt, task ID, or system fields.
• Submitting a guess when the video is broken, severely occluded, or unreadable.