Could we optimize the television news contextualizer by limiting it to just airtime seconds where there was a shot change, rather than analyzing each and every second of airtime? While intuitive, it turns out this approach misses a significant amount of material.
The Television News Contextualizer Through Shot Changes
