Standard vs Parallel Analysis: Which Should You Use?

Standard analysis

Standard analysis processes the whole video in a single pass. The model watches the footage from start to finish and returns every event it finds in one go, while a progress indicator takes you through downloading, uploading, reading the video's metadata and the analysis itself.

It's the simplest option and works best for short footage: highlights, a single play, a training drill, or a clip you want to check quickly. For these there's no benefit to splitting the video up, and a single pass keeps the whole context together. Standard results can be exported as an Excel spreadsheet.

Parallel analysis

Parallel analysis splits the video into short segments and analyzes several of them at once — three by default. As each segment finishes, its events are added to the timeline, so you can start reviewing the first half of a match while the second half is still being processed. A live view shows which segments are running, which are waiting and how many events have been found so far, and you can cancel the analysis at any point.

For a full match this is dramatically faster than a single pass. It's also more robust: if a segment fails, it's counted as failed and the rest of the analysis carries on and completes. Parallel results can be exported as JSON.

Setting up the segments

You can control the segments in one of two ways: by length, or by how many segments you want. By default each segment is 45 seconds long, and consecutive segments overlap by 2 seconds so an event that happens right at a boundary is seen in full.

It's tempting to think shorter segments must be faster. In practice they usually aren't. Only a limited number of segments run at the same time (three by default) and each one carries a small fixed overhead, so more, shorter segments often take longer overall than fewer, longer ones — and very short segments give the model less context around each event. For most footage, 45 to 60 seconds is the sweet spot.

The overlap has a side effect worth knowing about. Because consecutive segments share a couple of seconds, an event that falls inside that shared stretch can be picked up by both, and appear twice on the timeline. Keeping the overlap small keeps this rare, and when it does happen you can mark the extra entry as a duplicate or delete it.

Stream analysis: for matches that are live

Standard and parallel analysis work on recorded video. Stream analysis works on a match while it's being broadcast. You choose one of your organisation's live broadcasts and the sport, then set two things: the analysis window — how much of the most recent footage each step looks at, from 10 to 120 seconds (30 by default) — and the maximum run time, from 1 to 180 minutes (30 by default).

Once it starts, each new window of the broadcast is analyzed by the same engine as parallel analysis, and detections appear in a live feed, newest first. If one window fails, the analysis simply carries on with the next. It stops when the broadcast ends, when you press Stop, or when the maximum time is reached.

You can also choose to overlay the detections on the live stream, so viewers watching the broadcast see the events as they're found. And an organisation can have stream analysis start automatically whenever one of its broadcasts goes live.

Which to choose

Use standard analysis for short footage — highlights, clips and anything under about 20 minutes. Use parallel analysis for full matches, long training sessions and anything over 20 minutes, where the speed difference really shows. Use stream analysis when the match is happening now and you want events as they occur.

Same results, same tools

Whichever you choose, you get a timeline of timestamped events with teams and descriptions. You can click through them, filter them, mark each one as correct, wrong or duplicate, add your own events, or re-run the analysis with a custom prompt. Recorded-video analyses are saved to your history, labelled standard or parallel, so you can come back to them later.