RaceLabs · Internal Benchmark

We publish
our numbers.

We publish our internal June 2026 rider validation results, reported on identities excluded from training. Retrieval, group purity and group completeness are reported separately. Use the figures alongside a trial on your own event and a review of the resulting groups.

INTERNAL RIDER BENCHMARK · JUNE 2026 176 UNSEEN RIDER IDENTITIES · 6,505 PHOTOS RIDER MATCHING VALIDATION
0.991
Retrieval mAP
mixed benchmark
99.6%
Group purity
B³ precision · reported
86.3%
Group completeness
B³ recall · reported
6,505
photographs
in the mixed sample
Abstract

RaceLabs uses rider and vehicle signatures to group photographs across an event, including shots without a readable number. The June 2026 mixed benchmark covers 176 held-out identities and 6,505 photographs. It reports retrieval mAP 0.991, B³ precision 0.996 and B³ recall 0.863. These measure ranking, group purity and group completeness separately. They do not guarantee error-free or complete rider folders for every event; review and corrections remain part of delivery.

Benchmark card

The record.
The limits.

Source: RaceLabs’ published June 2026 rider-validation summary. This card preserves its reported aggregate figures and identifies what that public record does not specify. It is an internal evaluation, not an independently reproduced benchmark.

Date, version and domain

The published record is labelled June 2026. It does not state an exact evaluation day or a model/evaluation build identifier. Its reported domain is motorcycle rider matching, including a cross-championship split described as MotoGP / WSBK-class.

Sample and held-out identities

The June summary reports 176 identities and 6,505 photographs, with identities reported as excluded from training. The three splits list 25, 78 and 73 identities. Photograph counts per split and the query/gallery selection protocol are not published.

Metric and label reporting

The summary reports retrieval mAP, B³ precision, B³ recall and Adjusted Rand Index. It does not document the exact averaging or weighting for that June run, the label-review procedure, or how many images were excluded or left unassigned. Those missing counts limit any interpretation of overall coverage. A current evaluator definition does not establish how these earlier results were calculated.

What the figures cover

The results describe this rider sample. They do not measure end-to-end delivery, review time, processing volume, video, car or kart driver attribution, or athletics coverage. No independent replication package is supplied. Test a representative event and review its groups before delivery.

01 · Methodology

Reported on held-out
rider identities.

The published internal validation reports riders excluded from training. The definitions below explain how to read the metric names; the public summary does not specify the exact query selection, averaging or weighting used to calculate them.

mAP, mean Average Precision

Retrieval mAP, mean Average Precision, summarises how well photographs of the same rider rank ahead of other riders across retrieval queries. The mixed benchmark reports 0.991; the hard cross-championship split reports 0.998. Retrieval ranking is different from the completeness of a delivered rider folder.

Same-rider precision

B³ precision measures group purity: for a sample item, the share of its assigned group carrying the same identity label, aggregated across the evaluation. The mixed benchmark reports 0.996. This is a grouping metric, not a guarantee that 99.6% of delivered photographs are correctly assigned in every event. Review the groups for wrong-rider assignments.

What recall really measures

B³ recall measures group completeness: for a sample item, the share of its labelled identity group found in the same assigned group, aggregated across the evaluation. The mixed benchmark reports 0.863. A rider split across groups can lower recall and require a merge. Review both incomplete groups and wrong-rider assignments; high precision does not eliminate either check.

Grouping agreement and scope

Adjusted Rand Index (ARI) describes agreement between assigned groups and identity labels, adjusted for chance. The mixed result is 0.919. The listed splits describe the evaluated sample; they do not establish performance at every event.

02 · Results

Reported riders.
Reported splits.

These are internal reported values from the June 2026 summary, not an independently reproduced current run. No versioned run record links the figures to an exact evaluation build. Compare precision and recall together, with the missing excluded/unassigned counts in mind.

Internal reported June 2026 rider results; per-split photo counts and excluded/unassigned counts are not published
Validation split Unseen riders Photos mAP B³ precision B³ recall Cluster agreement (ARI)
Single event, clean conditionsbaseline split 25 Not published 1.000 1.000 1.000 1.000
Cross-championship, hardMotoGP / WSBK-class 78 Not published 0.998 0.993 0.961 0.981
Unseen new ridersreported new-rider split 73 Not published 0.982 0.996 0.722 0.819
Full mixed benchmarkInternalthe 176 listed identities 176 6,505 0.991 0.996 0.863 0.919

mAP mean Average Precision  ·  Precision / Recall B³ cluster metrics  ·  ARI Adjusted Rand Index, agreement with ground-truth grouping.
The new-rider split reports B³ recall of 0.722, compared with 0.863 on the mixed benchmark. Split groups may need merging, and wrong assignments may need correction. The scores do not establish that no photos are lost or misfiled in every event.

03 · Rider matching in practice

Different views.
Groups to review.

Number recognition reads visible digits; facial recognition compares visible facial features. RaceLabs uses rider and vehicle signatures to group an event. A readable number is not required, but obstruction, blur and similar equipment can still make a match uncertain. Judge the resulting groups on your own photographs.

Illustrative review situations, not measured comparisons with other products
SituationVisible informationWhat to review with RaceLabs
Number hidden in a packNumber may be obscuredRider signatures can help; inspect the group
Number angled awayNo readable digitsDifferent views can connect; inspect the match
Motion blurDigits or features may be blurredBlur can make a match uncertain
Practice without numbersNo number to readCan group without a readable number; review the result
Rain or mud over the plateNumber may be obscuredDifficult views still need review
Full-face helmetFace may be concealedA visible face is not required for rider grouping

A readable number can help name a group. Rider and vehicle signatures help connect photographs across the event. The number-reading and rider-signature guide explains that distinction in practice. Review uncertain matches before delivery.

04 · Review before publishing

Precision is what
the rider feels.

A rider expects the gallery to contain their photographs. B³ precision of 0.996 on the mixed benchmark measures group purity; it does not rule out errors in a customer gallery. Inspect wrong-rider assignments and split groups before publishing.

A practical review check

Inspect groups for photographs of another rider and for one rider split across several groups. The benchmark helps describe grouping quality; the event review determines what you are ready to publish.

Use the trackday grouping and review guide, then follow the editing, export and publication workflow to prepare the delivery.

The SensaPhoto customer story shows published rider galleries and photo/video offers. Those store examples illustrate delivery; they do not independently validate the benchmark or measure matching accuracy.

05 · Scope & limits

Use the scores.
Know their scope.

Review difficult matches

The Unclear bin provides a place to review uncertain images. Also inspect the rider groups, correct wrong assignments and merge split sets where appropriate. A change of motorcycle or an unusual view can make a match harder. The time required depends on the event.

Measured on our own domain

These are RaceLabs’ own published motorsport validation results. They cover the named splits and do not establish a result for every circuit, weather condition, rider or camera setup. Test a complete event and include review time when assessing the workflow.

06 · Questions

Accuracy, answered.

What does an mAP of 0.99+ actually mean?
mAP measures retrieval ranking: how well photographs of the same rider rank above photographs of other riders. The mixed benchmark reports 0.991, and the hard cross-championship split reports 0.998. This is separate from whether every rider folder is complete and correctly grouped.
Do you need the bib or number plate to be visible?
No. Rider and vehicle signatures help connect photographs across the event without a readable number. A visible number can help name a group. Difficult matches and split groups still need your review.
How does rider grouping relate to number recognition?
Number recognition needs readable digits for that part of the task. RaceLabs uses rider and vehicle signatures to group photographs across an event, including views without a readable number. Compare tools on a complete event and inspect both wrong assignments and split groups.
What does 99.6% same-rider precision mean for my delivery?
The mixed benchmark reports B³ precision of 0.996, or 99.6%. It measures group purity against benchmark labels. It is not a guarantee of correct assignment for 99.6% of the photographs in every customer delivery. Review the groups and correct errors before publishing.
Why is recall lower than precision on new riders?
B³ recall measures group completeness. One rider split across groups can lower recall even when those groups are relatively pure. The new-rider split reports recall of 0.722; the mixed benchmark reports 0.863. Review split sets for merging and inspect wrong assignments separately.
Is this an independently validated benchmark?
No. These are RaceLabs’ internal published results. The summary lists three rider splits and a mixed total of 176 held-out identities and 6,505 photographs. It does not supply a replication package or document the precise build, query protocol, weighting and label-review method. Read the benchmark card for the scope.
Do these figures cover cars, karts or athletics?
No. The published rider splits include a cross-championship sample described as MotoGP / WSBK-class. They do not establish car, kart or athletics outcomes. Evaluate each discipline on an appropriate event, including its review and access requirements.

Test it on your
hardest event.

Bring a complete event, including pack shots, difficult angles and practice sessions without numbers. Review the groups and corrections, then judge the workflow on your own photographs. Our guide to choosing motorcycle photo sorting software shows what to include in that evaluation.