Zion Time-Stamped Activity Descriptions
Semantic annotation pipeline for auditing and refining AI-generated captions for short-form video content. Experts review model-generated seed descriptions of hand actions captured from a robot's point of view, verify accuracy against video footage, and correct or approve annotations across nine edit categories. Annotations must accurately capture bimanual hand actions, object interactions, and spatial positioning without body movement descriptions or hallucinated details.