We publish
our numbers.
Sorting, solved. 30+ photographers · 20,000+ photos a minute · 100% group purity · mAP 0.996 · 5 live stores →
Our October 2026 benchmark on 175 held-out riders: 100% group purity, 0 photos in the wrong group, 96.5% of each rider in one group and 0.996 mAP.
B³ precision
wrong rider’s group
in one group
mixed benchmark
RaceLabs uses rider and vehicle signatures to group photographs across an event, including shots without a readable number. On 175 held-out riders and 6,504 photographs, the October 2026 benchmark reports 100% group purity: not one photo landed in the wrong rider’s group. Retrieval mAP is 0.996. Against our previous model, photographs in the wrong group fell from 16 to 0, and 96.5% of each rider’s photographs now land in one group (ARI 0.985). On our live servers, a 3,077-photograph event came back with no photograph in the wrong rider's group, checked photo by photo.
What was measured,
and on what.
Our October 2026 validation on motorcycle events: the sample, the metrics and their scope.
Date and domain
October 2026, motorcycle rider matching. It includes a MotoGP / WSBK-class cross-championship split, our hardest case: same circuits, similar bikes, similar leathers.
Sample and held-out identities
175 rider identities and 6,504 photographs, every identity held out of training.
Metrics reported
Retrieval mAP for ranking, B³ precision for group purity, B³ recall for completeness and Adjusted Rand Index for overall agreement.
What the figures cover
Rider grouping on motorcycle events. For speed, video, cars, karts or athletics, run a complete event of your own.
Reported on held-out
rider identities.
Every rider below was held out of training. Four metrics, each answering a different question.
mAP, mean Average Precision
How well photographs of the same rider rank ahead of everyone else: 0.996.
Same-rider precision
Group purity: the share of a rider’s group that really is that rider. 1.000: not one photo in the wrong group. The number a rider feels.
What recall really measures
Group completeness: the share of a rider’s photographs that land together. 0.965.
Grouping agreement and scope
Adjusted Rand Index: how closely the whole grouping matches the ground truth, chance removed. 0.985.
Previous model.
RaceLabs today.
All 175 held-out riders, 6,504 photographs.
| Model | mAP | B³ precision | B³ recall | ARI | Photos in the wrong group |
|---|---|---|---|---|---|
| Previous model | 0.993 | 0.997 | 0.853 | 0.913 | 16 |
| New model | 0.996 | 1.000 | 0.965 | 0.985 | 0 |
mAP mean Average Precision · Precision / Recall B³ cluster metrics · ARI Adjusted Rand Index, agreement with ground-truth grouping.
Across all 175 riders the model had never seen, precision is a perfect 1.000: not one photo in the wrong group.
No number,
still sorted.
Number readers need the digits. RaceLabs groups by rider and vehicle signatures, so it still places the frames that hide the number.
| Situation | Reading the number | RaceLabs |
|---|---|---|
| Number hidden in a pack | No digits to read | Grouped by rider signature |
| Number angled away | No digits to read | Connected across views of the same rider |
| Motion blur | Digits unreadable | Placed when the signature holds, sent to Unclear when it does not |
| Practice without numbers | Nothing to read | Grouped without a number at all |
| Rain or mud over the plate | No digits to read | Grouped by rider signature |
| Full-face helmet | No face, no digits | A visible face is not required |
A readable number still names the group once the photographs are together. The number-reading and rider-signature guide explains the difference.
Precision is what
the rider feels.
A rider opening their gallery notices one thing first: a photograph of somebody else. B³ precision measures exactly that. Ours is 1.000: not one photo of somebody else in any rider’s group.
A practical review check
Check two things before publishing: another rider inside a group, and one rider split across several. The first did not happen once in 6,504 photos; the second is rare: 96.5% of each rider’s photos land in one group, and the Rear Finder offers any rear-only set to its most likely rider. In photographers’ own events: 100% of rear-view sets placed, 97% on the very first suggestion.
Use the trackday grouping and review guide, then the editing, export and publication workflow.
The SensaPhoto customer story shows this workflow ending in published rider galleries.
Use the scores.
Know their scope.
Review difficult matches
Uncertain frames go to the Unclear bin, never to the wrong rider. That covers a mid-event bike change or an unusual angle.
Measured on our own domain
The method and all four metrics are published. The best test is a complete event of your own: include the review time when you compare.
Accuracy, answered.
What does an mAP of 0.99+ actually mean?
Do you need the number plate to be visible?
What does 100% group purity mean for my delivery?
How was the benchmark run?
Do these figures cover cars, karts or athletics?
Test it on your
hardest event.
Bring a complete event with pack shots, hard angles and practice without numbers, and judge the groups on your own photos. Our guide to choosing motorcycle photo sorting software shows what to test.