Video is a ubiquitous medium. From TV and movies to social media and mobile platforms, its applications are numerous in the entertainment sector. The ability to ‘describe’ what’s happening in a video, without human intervention is the eventual goal of an AI/ML engine. The usual definition of Metadata is ‘data about data’. However, in the context of machine learning, Metadata are the ‘content descriptors’ of video/audio/textual data elements extracted from a video. Video metadata extraction is the process of auto-identifying video content.
An emerging trend is the application of AI technology for TV advertising. In this Best Practices Guide, we discuss the unique challenges in applying machine learning to carrier-class video advertising. To illustrate the point, the discussion is focused on a specific use case that is common to all ad supported TV services.
The selected use case is Ad Ingest Quality Control (QC). TV commercials are subjected to various rules and regulations. For example, ads containing specific content (e.g. Alcohol, firearms) are barred from airing during certain TV programs. Identifying these categories might pose a challenge to a machine learning tool, as off-the-shelf products are more oriented towards facial recognition. That is to be expected perhaps, as the video ML products were primarily intended for surveillance and sports applications. However, by judiciously combining metadata from multiple data streams, ML based analysis can be enhanced.
This best practices guide consists of two parts. The solution description section contains recommendations based on AI/ML WG member experiences and the lessons learned. The requirements section defines features a machine learning tool would need to perform for specified tasks.
Machine learning based video analysis is a burgeoning field. As the technology matures, its applications in broadband industry will be far and wide. Network operators and service providers will find the guidelines useful in solution development. Vendor partners who are developing carrier-class machine learning solutions might embody these in product specifications.
The intended audience are the machine learning practitioners in cable-telecom space.