Making sense of footage rather than generating it: summarising a recording, finding the moment something happened, describing what occurred, checking whether a procedure was followed. Given how much video organisations hold and how little of it anyone watches, this is quietly one of the more valuable applications.
Technical view
Approaches sample frames and process them with a vision language model, with the sampling strategy dominating both cost and quality, since dense sampling is expensive and sparse sampling misses events. Long videos require hierarchical summarisation or retrieval over segments. Audio tracks add substantial signal and are often ignored. Temporal reasoning, ordering and causality across a long recording remains markedly weaker than single-frame understanding.
Related
Part of AI Terminology Index, section: Modalities and Applications.