← FishCoin

Measure a fish

Upload a few seconds of video with a FishCoin token lying next to the fish, and we will measure it against the token. Free, and no account.

Read this before you upload — we keep what you send.

This is not a calculator that forgets you. The clip you upload, and the location recorded inside it if there is one, are kept permanently and become part of the data FishCoin is built on. We use captures to improve the measuring software. Measurements and positions may in future be supplied to fishery-management and research bodies — that is not happening today, and the privacy policy will say so before it starts.

That is the exchange: you get the measurement for nothing, we get the capture. If you would rather not, do not upload — there is no reduced version of this page that measures without keeping.

Full detail in the privacy policy and the terms of service, section 4 of which is the licence you grant us in what you upload.

What to send

A single photograph will not work, and that is measured rather than a preference. The measuring system judges a capture on agreement between frames, and one frame cannot agree with anything — every still we tested came back "needs review" and none could ever be published, however good the photograph. We would rather say so here than take your upload and hand you a shrug.

Upload

What the number is worth

It is an estimate from a camera, and we do not guarantee it. Against tape measurements of the same fish, catches good enough for us to publish averaged about 12% away from the tape, and only about two in three were within 15%. The system also reads roughly 7% long overall and we have not corrected for it.

Do not use this to decide whether a fish is legal to keep. The error above is wider than the margin most size limits leave you, and a reading from a phone is not a defence. Measure a fish you intend to keep with a ruler or a brag mat.
And you are a harder case than anything we have measured. Every capture this system has been tested against was shot by an angler using the FishCoin app, who knew how it wants the fish laid out. Nobody has measured what it does with a clip from someone who has never seen the app. Expect it to do worse than the figures above, not better.

Identification — what we will and will not tell you

We have a species identifier. This page does not show you its answer, and that is a decision made from measurements rather than an unfinished feature.

It does well on curated reference photographs. That does not survive contact with real catches. On photographs actually sent in from boats and beaches it named the right broad group — finfish, shark and ray, crustacean, cephalopod — 58% of the time. Answering “a finfish” every single time, without looking at the picture at all, would have been right 96% of the time.

There is a way to tell when it is more likely to be right. It still is not enough — and that, rather than the model being noise, is the finding.

The sensible proposal is to answer only when the model is confident: when every frame of a clip agrees. That confidence is not meaningless. On real catches it does separate the answers it gets right from the ones it gets wrong, well beyond chance. So anyone who proposes thresholding on it is not wrong that the signal is there.

The problem is what is left behind the threshold. Answering only when every frame agrees lifts the model to 77% — and on those same clips the constant “finfish” scores 94%. We tried five such rules, written down in advance so we could not go fishing for a flattering one. Not one of them beat answering “finfish” every time.

On the strictest of those rules the two are closer than they used to be, and it is worth being exact about what moved: the model is still behind. What weakened is our confidence that the constant is reliably ahead on that smaller sample — not any evidence that the model has caught up. Closer is not nearly there.

And one thing about those real catches, which makes all of the above a best case rather than an expected one: they were taken by anglers using the FishCoin app, who knew how it wants a fish laid out. A stranger’s single upload here is a harder case than any of them.

These figures changed on 19 August, and not because anything here was wrong. We replaced the identifier itself, so the older numbers on this page were not stale readings of the current system — they were accurate readings of a system we no longer run. It is a distinction worth making, because a page can be perfectly faithful to its evidence and still end up describing something that no longer exists.

So there is no species, no group and no confidence score on this page, because none of them would mean what a reader would reasonably take them to mean. If that changes, it will be because the measurement changed, and this section will say so.