Activity classification
Also known as: clip categorisation, activity tagging
Sorting a whole clip or upload into one activity, such as a surf session or a park lap, rather than naming the movements inside it.
Updated
What it means
Activity classification answers a coarse question about a recording: what is going on here. A clip is filed as a surf session, a park lap, a dive or a family afternoon, and that single label travels with it through the rest of the pipeline.
It sits one level above Action recognition, which names individual movements inside a clip. Classification is far more stable than recognition, because a long recording almost always belongs to one activity even when the moments inside it vary wildly from second to second.
Because the answer is one label for many frames, it is among the cheapest passes a Computer vision chain can run. Most systems do it first, before anything expensive looks at the footage in any detail, so that later stages know what they are dealing with.
How it works in practice
The label mainly sets expectations for later stages. A surf clip is mostly waiting, so the search narrows to a few promising seconds in every minute. A moto clip is mostly movement, so the problem inverts into choosing between many good options.
The same label decides which template the cut starts from and how aggressive Smart trimming should be. A dive wants long takes and slow pacing; a skatepark session wants short ones and a faster rhythm, and neither setting suits the other.
It also narrows comparisons. Two recordings of different activities are never near-copies of each other, so classification shrinks the work Duplicate shot detection has to do across a large mixed upload from several days.
What to watch out for
The long continuous recording is the weak case. Forty minutes that start in a car park, cover three lift laps and end at a table have no single correct answer, so classification has to run on segments rather than on whole files.
Running on segments sacrifices some of the stability that made the pass attractive, and the boundaries between segments are themselves a guess. A stretch that spans the walk back and the next run can end up labelled as neither.
A wrong label is expensive because everything downstream inherits it. A dive filed as snorkelling gets the wrong pacing and the wrong Shot scoring thresholds, and the result looks oddly edited with no obvious fault in the footage.
How RawClip handles it
Every upload is classified before the search for moments begins, which is why a surf session and a bike park session come back cut to different rhythms from the same system, without anybody choosing a style by hand or filling in a form about the sport.