Temporal segmentation
Also known as: time segmentation, clip splitting
Dividing footage along the timeline into stretches that can be judged separately, which is what makes a long recording analysable at all.
Updated
What it means
Temporal segmentation cuts a recording into spans of time rather than into shots. Where Shot change detection looks for boundaries in the picture, this looks for boundaries in what is happening.
A single uninterrupted shot can contain an approach, a jump and a long roll out, and those are three different things to score even though the camera never cut between them.
Segmenting first is what lets everything downstream work on comparable units. A score attached to forty minutes is useless; a score attached to each of two hundred stretches is a ranking.
How it works in practice
The simplest version is a fixed window with overlap, which guarantees coverage and never misses a boundary because it never looks for one. It is crude and remarkably hard to beat.
Signal driven segmentation does better where the signals are reliable. Motion onsets and audio events mark natural edges, so Motion detection usually supplies the candidate boundaries.
Overlap matters more than precision. A moment straddling two segments is scored twice and the better score wins, which is far safer than a boundary that splits it and loses both halves of it.
What to watch out for
Too fine a division destroys context. Two second segments contain no approach and no outcome, so everything scores alike and the ranking stops distinguishing anything.
Too coarse a division hides the moment. A ninety second segment containing one brilliant second averages out to unremarkable, and nothing ever looks closely enough to notice.
Segment boundaries also become edit points by default, which is rarely right. A cut belongs slightly before the action, so the marks need adjusting afterwards by Auto-cut rather than being used exactly as the segmentation left them. Treating an analysis boundary as an editorial one is a common and avoidable mistake.
How RawClip handles it
Uploads are divided into overlapping stretches before anything is scored, so a moment that happens to fall near a boundary is still judged whole. The boundaries are used for analysis only, and the actual cut points are placed separately against motion and sound.