Abstract
We present a deployed AI-assisted video annotation system that achieved an 83% reduction in production annotation time through structured formalization of expert cinematographic knowledge. Evaluated across 12 production projects over 3 months, our system reduced annotation requirements from 150–200 person-hours to 25–30 person-hours while improving annotation accuracy (91% vs 89% F1) and consistency (94% vs 82% terminology standardization).
Our core contribution is a methodology for encoding tacit director expertise into JSON-formatted schemas that enable Google Gemini to serve as an intelligent annotation assistant. The system structures cinematographic conventions—shot types, camera movements, composition principles—into machine-readable specifications that guide AI analysis and enable iterative human-AI validation loops.
Deployed in multilingual production environments supporting nine languages, the system demonstrates 95% translation accuracy for technical terminology and 91% cross-lingual semantic consistency. Statistical analysis across projects shows significant improvements in temporal precision, reduction in revision cycles, and decreased multilingual communication overhead.
Contact: sedat@hagiaproject.com