Video annotator: a framework for efficiently building video classifiers

The Netflix Technology Blog introduces the Video Annotator (VA) framework, aiming to efficiently build video classifiers using vision-language models and active learning techniques. The framework addresses challenges in traditional machine learning model training by involving domain experts directly in the annotation process. VA leverages active learning and zero-shot capabilities of large models to guide users in focusing on harder examples, enhancing sample efficiency, and reducing costs. By integrating model building into the annotation process, VA allows for user validation before deployment, fostering trust and ownership. The framework supports continuous annotation, enabling rapid model deployment, quality monitoring, and quick resolution of edge cases. Experiments show VA outperforms baseline methods in creating high-quality video classifiers, demonstrating an 8.3 point improvement in Average Precision across various video understanding tasks. VA empowers domain experts to make improvements independently, enhancing efficiency and trust in the system.