What is logo detection?

How it works

There are three broad approaches. A trained detector learns a fixed list of brands from many labeled examples and draws a box around each one it sees; it is accurate for those brands and blind to the rest. An open-vocabulary model matches images against text, so you can type a brand name it was never specifically trained on and still get results for well-known marks. Reference matching takes a sample image of the logo and looks for regions in each frame that resemble it, which works for a local business or a new rebrand no model has seen.

In video, detection usually runs on sampled frames rather than every frame, and nearby hits are merged into time ranges: the logo was visible from 12:04 to 12:31.

An example

A creator's sponsor asks for proof of how long its logo was on screen across a season of videos. Running logo detection over the episodes produces a list of time ranges per video. The creator checks a few, adds them up and sends the report, instead of scrubbing through hours of footage with a stopwatch.

The same run answers other questions too: which clips show a competitor's logo that a new sponsor would object to, or which archive footage contains a mark that needs to be blurred or cleared before a client can air it. In each case the value is the same, turning “somewhere in these hours” into a list of time ranges a person can check.

Why logos are hard

• Small and far away. A logo on a jersey across a field may cover a few dozen pixels.
• Distorted. Logos on bottles, flags and shirts bend, fold and sit at angles.
• Blurred or cut off. Motion blur and partial views are the norm in real footage.
• Many versions. Brands have wordmarks, icons, one-color and inverted variants, and they redesign.
• Text overlap. A wordmark is also text, so OCR may find it, while an icon-only logo needs visual matching.

In MediaFind

Brand-logo detection is a Pro feature in MediaFind and runs on your computer. You can find where a brand appears by typing its name, for example “Nike,” or by uploading a sample logo image, which helps with smaller brands. Results open at the moment the logo is on screen. Wordmarks and signs are also reachable through on-screen text search, which is free. For a walkthrough, read how to find a logo in video; for context, see MediaFind for YouTubers.

Frequently asked questions

Can logo detection find a logo it was never trained on?

Sometimes. Open-vocabulary models can find well-known brands by name, and matching against an uploaded sample image can find logos no model has seen.

Is logo detection the same as OCR?

No. OCR reads text, so it can catch wordmarks. Logo detection also finds symbols and icons that contain no letters.

What is logo detection used for?

Measuring sponsor exposure, checking brand safety, clearing footage for unlicensed marks before publishing, and finding clips for brand work.