Video summarization
Also known as: automatic summarisation, video digest
Reducing a long recording to a much shorter one that still conveys what happened, which is the academic name for making a highlight.
Updated
What it means
Video summarization is the research framing of the same problem a highlight solves. Given a long input, produce a short output that preserves what mattered and discards the rest.
The literature usually measures success against human chosen segments, which is a reasonable proxy and also the source of most disagreement. Two people summarising one session rarely pick the same seconds.
The distinction from a finished highlight is pacing and music. A summary is a correct selection; an edit is a correct selection arranged to be watchable, which is what Storyboard generation adds.
How it works in practice
Most approaches score stretches and take the top ones subject to a length budget, which is exactly what a Highlight detection model does in production.
Coverage constraints are the common refinement. Requiring that the summary touch every part of the recording stops it collapsing onto the single most eventful minute and leaving the rest of a session unrepresented.
Diversity constraints do the other half, forcing the selection to avoid repeats so the result is not four views of one moment. Both are applied during selection rather than afterwards, because filtering a finished selection means rebuilding it.
What to watch out for
Optimising for agreement with human labels rewards the average taste and penalises a distinctive one, so a technically excellent summary can feel characterless.
Length budgets also distort the result. A fixed thirty seconds forces the same treatment onto a quiet afternoon and an exceptional session, when the honest answer is two different lengths, which is what Smart trimming against a flexible target delivers.
Summaries built without grouping repeats look worst of all. Every selected moment is individually justified and the whole reads as a loop, which is the Near-duplicate clip problem in its purest form.
How RawClip handles it
A highlight is built as a selection with coverage and variety constraints rather than simply the top scoring seconds, and the target length flexes with how much the footage actually contains. A quiet session comes back shorter rather than padded out with weaker material to fill a fixed duration nobody asked for.