Accuracy
How we measure accuracy, so you can hold us to it.
A count is only useful if you know how far to trust it. This page explains how Fincount measures its own error, and the targets every model version has to meet before it is used.
The method
Four rules for the test.
- 01
Two hand counts of every basin
Rule 1
Two people count each evaluation basin by hand, separately. Their agreement is reported next to Fincount's error, because a count can only be as good as the reference it is checked against. - 02
Hatcheries and days the model never saw
Rule 2
The evaluation set comes from hatcheries and days that are not in the training data. It is frozen before any model is tested on it and never used to tune a model. - 03
At least 300 basins
Rule 3
Spread across the three size groups, across densities up to the limit and one level above it, in white and blue basins, in shade and in open sky, on a stand and handheld. - 04
Error bars on the error
Rule 4
Every measure is reported overall and by size group, density, basin colour and light, with a 95% interval on the measure itself, so a claim like "within 3% for 95% of basins" comes with its uncertainty.
The targets
What each model version has to meet.
These are the targets Fincount is held to. A model is only released when it meets them on the evaluation set, and the report is stored with the model version.
Count of one basin
Target: within 3% for 95% of basins.
Count of one sale
Target: within 2% across several basins.
Basins in the test
At least 300, spread across sizes, densities and light.
The range itself
The 95% range holds the true count for about 95% of basins.
| Measure | Target |
|---|---|
| Count of one basin | Within 3% for 95% of basins, never worse than 5%, up to the supported density |
| Count of one sale | Within 2% across several basins |
| Length of a fish | Within 3 mm or 5%, whichever is larger |
| Average weight of a batch | Within 10% |
| Size group of a fish | 90% of fish in the right group |
| First count on the phone | Under 10 seconds for a 10 second video, with no signal |
| Final count in the cloud | Under 2 minutes after upload |
| Capture checks | Reject 95% of unusable videos and fewer than 5% of good ones |
| The range itself | The 95% range holds the true count for about 95% of basins |
What a range means
Margin of error, in plain words.
Every count comes with a range, for example between 482 and 508. The range is the margin of error: Fincount is built so that the true count falls inside it for about 95 of every 100 basins. The range is set from basins that people counted by hand, and it is worked out again for every new model version, because a range that was right for one model can be wrong for the next.
Crowded basins get wider ranges, and past the crowding limit Fincount asks you to split the basin. When two counts of the same sale are compared, a difference is only a dispute when it is bigger than the combined range. Nothing on this page is a promise about one particular basin. It describes how the test works.
An example range
Between
and
The true count falls inside the range for about 95 of every 100 basins.
Check it yourself
You do not have to take our word for it.
Recount any basin by hand and compare. Upload a basin video on the try page and see every dot. Ask us for the evaluation report of the model version you use.
Count your next basin with Fincount.
Film a basin, get the count, and send your buyer a count they can check. The first hatcheries get 30 days free.