Confidence score
Also known as: certainty score, detection confidence
A number a model attaches to its own answer, saying how sure it is, used to decide which detections to act on and which to ignore.
Updated
What it means
A confidence score is the model saying how much it believes its own output. A face is reported at nine tenths certainty, a bicycle at a third, and the pipeline decides what to do with each of them.
Without it every detection would carry equal weight, which would mean acting on guesses as readily as on plain facts. The score is what lets a system be decisive about the obvious and cautious about the rest.
The number is not a probability in any strict sense. It is a comparable ranking produced by one model on one kind of input, which is enough to sort and threshold but not enough to quote as a chance of being right.
How it works in practice
Thresholds are where the score earns its keep. A high bar means only certain detections survive, which suits anything irreversible such as discarding footage, while a low bar suits gathering candidates something else will filter.
That is why Shot scoring can afford to be generous where a delete step cannot. The same detector feeds both, and only the threshold differs between them.
Scores are also the natural trigger for review. Anything sitting in the uncertain middle is exactly what Human-in-the-loop editing should look at, while the confident ends of the range need nobody at all.
What to watch out for
Models are often confidently wrong on material unlike their training data. Dim underwater footage or a heavily fisheye frame can produce high scores for detections that are simply not there.
That is one route to a Hallucination nothing flags, because filtering by certainty cannot remove an error the model is certain about. Cross checking against measurable signals is the only real defence.
Comparing scores across models is another trap. Seven tenths from one detector and seven tenths from another mean different things, so every threshold has to be tuned per model rather than carried over from the last one.
How RawClip handles it
Detections below a working threshold are treated as absent rather than uncertain, so a shaky guess about what is in a clip never becomes a reason to keep it or drop it. Only the confident signals reach the scoring stage.