Segment scoring
Also known as: window scoring, interval scoring
Rating time windows inside a long recording rather than whole clips, which is how a continuous file becomes a set of candidates.
Updated
What it means
Scoring a clip assumes clips exist. A mounted camera produces one forty-minute file, so the unit being judged has to be a window of time. Nothing in the recording marks where a run began or ended.
Segments are scored on the same signals as clips. Motion, composition, sharpness and audio, measured across a few seconds at a time and compared against their neighbours in the same file.
It differs from Shot scoring in the unit, not the method. One ranks things somebody already separated; the other has to separate them before it can rank anything.
How it works in practice
Window length is the main parameter. Too short and a movement is split across segments; too long and a good moment is diluted by the approach around it. A few seconds is the usual compromise.
Overlapping windows avoid the boundary problem. A peak landing exactly between two segments scores badly in both, which is why windows are stepped rather than adjacent. The result is more of them to compare.
Scene boundaries improve it where they exist. Shot change detection gives natural edges to score between, and Temporal segmentation is what turns a continuous file into those units.
What to watch out for
Scores are only comparable within a session. A quiet morning and a busy afternoon produce different distributions, so an absolute Inclusion threshold behaves differently on each card.
Repetition survives segment scoring. Four attempts at one feature produce four similar high scores, which is why Duplicate shot detection runs after the ranking rather than before it.
Context is invisible at this scale. A segment is judged on its own few seconds, so a moment that only works because of what preceded it scores no better than one that stands on its own.
How RawClip handles it
Long continuous recordings are handled this way: the file is divided into windows, each is scored, and the strongest are kept as candidates before near duplicates are collapsed. That is why a single long file produces a usable selection rather than one enormous clip with the interesting parts buried somewhere inside it.