How AI logo detection works, and where it fails

Logo detection sounds like magic and is mostly arithmetic. Understanding where it works and where it does not is what separates a credible report from an impressive-looking one.

What happens to an image

1. Reference marks for each partner are loaded once — ideally several variants, since brands appear in different colourways.
2. Every image the club publishes is scanned against those references.
3. Each candidate match returns a bounding box and a confidence score between 0 and 1.
4. Anything above threshold is recorded. Anything below goes to human review.
5. Recorded detections are scored for prominence — share of frame, position, occlusion.
6. Prominence is multiplied by that post's reported reach to give an exposure figure.

What confidence actually means

It is the model's certainty that a region matches a reference, not a measure of how valuable the exposure was. A 0.94 on a tiny background logo is worth far less than a 0.71 on a banner filling a quarter of the frame.

Confidence decides whether something is counted. Prominence decides what it is worth. Conflating the two produces nonsense.

What it is very good at

• Designed graphics — matchday cards, score graphics, team announcements. Near-perfect, and this is usually the majority of a club's output.
• Perimeter boards — flat, high-contrast, consistent. Reliable.
• Posed photography — squad photos, presentations, sponsor activations.

What it is bad at

• Small chest logos in fast action photography
• Heavily occluded marks
• Logos rendered in unusual colourways not supplied as references
• Video — possible, but far more expensive per unit of content

Why the failures should be published

A report claiming to have found every instance is a report a media agency will test and break. One stating that 92% of content was analysed, with the remainder listed and the reason given, survives scrutiny.

This is why our reports carry a coverage statement, and why anything below threshold goes to a verification queue rather than quietly into a total. The rest is on sponsorship measurement.

What the pipeline does to a single image

Step by step, with what can go wrong at each stage:

Most of the failure modes are decisions rather than technology limits. Where the threshold sits, whether clutter is scored, how much platform-reported reach is discounted — these are choices, and a club should know what they were set to.

Real numbers from testing

We ran detection across match photography and designed graphics. The results were consistent and worth publishing.

That last row is the honest limit. A chest logo at realistic size in action photography returns almost nothing. Anyone selling reliable jersey detection in match action is overselling, and it is worth asking a competitor to demonstrate it on your own photographs rather than their showreel.

Why the majority case still works

The limitation matters less than it sounds, because of what clubs actually publish. Across a typical club's output, roughly three-fifths is designed graphics — matchday cards, team announcements, score graphics, quote cards. Detection on those is essentially perfect, because the marks are placed at known size on a flat background.

Add perimeter boards in photography and posed shots, and the reliably measurable portion covers the substantial majority of published content. The unmeasurable remainder is action photography of kit, which is exactly the portion a coverage statement should disclose.

False positives, and why they matter more than misses

A missed detection understates a partner's exposure, which is embarrassing. A false positive overstates it, which is worse — because when the sponsor's team spots one, every other number becomes suspect.

Guards worth having:

• A maximum area filter, so a whole-frame artefact cannot be recorded as a detection
• Human review for anything between the noise floor and the confidence threshold
• Spot-checking a random sample each month against the source images
• Storing every box and score, so any figure can be traced back and audited

That last one is the difference between a number and a claim. If a report says a mark appeared 1,724 times, somebody should be able to ask to see them.

What to ask a vendor about their detection

• What confidence threshold do you use, and what happens below it?
• Does prominence scoring account for competing brands in frame?
• How much do you discount platform-reported reach, and why?
• Can you show me detections on my own photographs before I sign?
• Does the report state what proportion of content was analysed?
• Can I see the bounding boxes behind any figure in the report?

A vendor comfortable answering all six is measuring. One who deflects on thresholds and coverage is producing numbers rather than measurements, and the difference will surface in front of a sponsor rather than in front of you.

Read next