Threshold tuning
Also known as: cutoff tuning, sensitivity tuning
Choosing the cutoffs that turn scores into decisions, which is where most of the visible behaviour of an automatic editor actually lives.
Updated
What it means
Threshold tuning sets the numbers that convert continuous scores into yes or no. A detector reports a certainty, a scorer reports a rank, and a threshold decides what happens next.
Those cutoffs shape the output more than the models do. The same scores with a higher bar produce a short, punchy Final cut and with a lower one a long, contextual one from identical footage.
It is also where a Confidence score stops being informative and starts being consequential, because above the line a detection is treated as fact and below it as absent.
How it works in practice
Different decisions want different bars. Discarding footage is irreversible and deserves a high bar, while gathering candidates for later filtering can afford a generous one.
Tuning is done against a fixed set of real uploads rather than by eye, because every change looks like an improvement on whichever clip happened to be open at the time. A few dozen clips with known answers turn the work into a measurement.
Thresholds also need to be per activity. Brightness limits that suit forest footage condemn a snow session, which is why Activity classification runs before the bars are applied.
What to watch out for
A single global bar is the usual mistake. One number across cameras, light conditions and sports is wrong somewhere by construction, and the failures look like model errors rather than configuration.
Tuning on aggregate accuracy hides the cases that matter. A bar that is right on average can be wrong on every underwater clip, and nothing in the overall figure says so.
Thresholds also silently expire. A model update shifts the distribution of scores, so yesterday bar sits in a different place today, which shows up as a change in Smart trimming behaviour nobody asked for.
How RawClip handles it
The cutoffs that decide what counts as a detection and what reaches the cut are tuned against a fixed set of real uploads and set separately per activity, so a snow session and a forest ride are not judged against the same brightness bar.