Independent review · Any vendor
Find the label errors before your model learns them.
We review sports labels from any vendor or in-house team against your spec. Every error is confirmed by a second specialist, located by file and frame, and counted.
BOXES · IDS · KEYPOINTS · MASKS · EVENTS
REVIEWED
8
FLAGGED
4
CONFIRMED BY
2nd specialist
Illustrative simulation with planted errors. Real reviews follow your spec.
WHEN · 01
Before a training run
Find out what the model is about to learn before you spend the compute.
WHEN · 02
When a model plateaus
Check whether the ceiling is in the architecture or in the labels.
WHEN · 03
Before switching vendors
Get evidence instead of a feeling, on the same footage and the same spec.
WHEN · 04
Before you buy a dataset
Know the error rate of a dataset before it becomes your problem.
What we check
The 6 errors that cost models most.
A frame can look perfect and still be wrong over time. These are the error types we look for, and how each one is counted.
Identity swaps
Track IDs that jump between players after a screen, a pile-up or a camera cut.
- EXAMPLE
- ID 7 follows the wrong defender for 40 frames after two players cross.
- HOW WE MEASURE IT
- Every ID followed from first to last frame; swaps counted at each cut and collision.
What you get
A report, not a vibe.
Error rates by type and class from a sample designed to represent the whole dataset, every confirmed error with its file and frame, and a short list of what to fix first.
If your guidelines are missing or vague, you also get the written rules we reviewed against.
QUALITY REVIEW · EXAMPLE REPORT
Football events · Vendor B
SHARE OF CONFIRMED ERRORS BY TYPE
- Identity swaps34%
- Missing labels22%
- Wrong event18%
- Late / early events14%
- Loose geometry8%
- Schema drift4%
FIX FIRST
- 1. Re-link IDs after screens. Most swaps start within two seconds of a screen.
- 2. Agree a shot timestamp rule. The spec does not say strike or landing.
- 3. Add far-side players to the inclusion rule. They are skipped when small.
EVERY ERROR LOCATED
| match_014.json | f 18,402 | Identity swap | ID 7 ↔ ID 11 after screen |
| match_014.json | f 18,955 | Wrong event | Blocked shot tagged as pass |
| match_021.json | f 6,120 | Missing label | Far-side full-back unboxed |
| match_022.json | f 41,388 | Late event | Shot +6 frames |
Example layout with invented figures. Your report uses your spec, your classes and your data.
How it works
Labels in. Evidence out.
- 01
Share labels and spec
The labels, the footage they belong to, and the guidelines they were made to. No guidelines? We write down the rules we review against first.
- 02
Agree a sample plan
Which matches, classes and hard cases to review, and how many, so the error rates mean something for the whole dataset.
- 03
Review in two passes
A sport specialist reviews the sample against the spec, and a second reviewer checks every flagged item before it is counted.
- 04
Get the report
Error rates by type and class, every error located by file and frame, and a ranked list of what to fix first.
Reviewed by people who know the sport
A generic reviewer passes what it does not know.
A box can be tight and still be on the wrong player. Our reviewers know the sport they check, so they catch the logic errors a pixel check cannot.
Football
Basketball
Tennis
Ice hockey
Plus cricket, rugby, American football, baseball, golf and more. Football tagging · Basketball annotation
Why you can trust the numbers
No single opinion counts as an error.
A review is only useful if its error count holds up when your vendor pushes back. So every flag is confirmed by a second specialist, and the report is audited before it reaches you.
The metrics we use, explainedPrimary annotator
Reviews the sample against your spec and flags every suspected error with file and frame.
Specialist reviewer
A second sport specialist confirms or rejects each flag, so the report only counts real errors.
Senior auditor
Checks the numbers, the sample plan and the fix list, and signs off the report.
Pricing
Quoted per dataset. Start with a sample.
PER DATASET
Priced by sample and label type
The quote depends on how much we review and what kind of labels they are. Correcting the errors we find is optional and quoted separately.
FREE PILOT
See what we find first
New clients start with a free pilot review on a small sample of their labels, so you can judge the findings before paying for anything.
Questions
What ML teams ask us.
Which label formats can you review?
COCO, YOLO, MOT, CVAT and Label Studio exports, JSON or CSV event logs, keypoints and segmentation masks. If your format is custom, send the schema and a sample file.
What if we don't have written annotation guidelines?
Then the first step is writing down the rules we review against, from your model's needs and a look at the labels. You get that document too, and it is often as useful as the review.
How large a sample do you review?
It depends on the size of the dataset and how precise you need the error rates to be. We propose a sample plan before starting, weighted towards the hard cases where errors hide.
Will you fix the errors you find?
If you want. The review stands on its own, and correcting flagged items or relabelling a batch is quoted separately.
Can you review labels from several vendors at once?
Yes. We review each vendor's labels against the same spec and the same sample design, so the results can be compared directly.
Do you sign NDAs?
Yes. We sign a mutual NDA before any labels or footage are shared, and we are happy to use your template.
Independent review
Know what is in your data before you train on it.
Send a sample of the labels you are worried about. We will tell you what is wrong with them, and how often.