What this is
This site shows what an underwater camera in a pond sees: the fish that swim past, when they tend to show up, and the clips worth watching again. A program called PondWatcher runs on a PC cabled to the camera. It records a clip whenever something moves, and tries to name what it was. This site only displays what PondWatcher publishes. Nobody can change anything from here.
How it works
- 1 Camera An underwater camera films the pond around the clock.
- 2 Motion boxes A program on a PC beside it draws a box around anything that moves.
- 3 First look A fast, cheap outside model says how likely it is that a box holds a fish or a turtle.
- 4 Claude If there may be an animal, Claude is asked to name it.
- 5 Local model A small model on the PC scores every box too, and can call a fish the others gave no name.
- 6 The owner The owner goes through the clips and marks what is a fish, what is faint, and what is not a fish.
- 7 Nightly retrain Each night the local model is trained again on those marks.
The first job is noticing that something moved. The detector looks for sharp movement and for large faint shapes, and draws a box around each. Most boxes are not fish, so each one worth a look is cut out of the picture and checked.
The check happens in two steps. A fast outside model takes the first look and answers in about a quarter of a second: fish, maybe a fish, or not a fish. Only when there may be a fish or a turtle is Claude asked for a name. If Claude cannot say and the first look thought a fish was likely, a larger Claude model looks at the same picture. A fish that got no name is looked at again while it stays in view, up to four times. Every look at an outside model costs a fraction of a cent, and each service has its own daily spending limit. The AI page shows what was asked and what it cost.
A small model trained on this pond's own clips runs on the PC and scores every box as well. It costs nothing to run. It cannot tell one type of fish from another, but where the outside models gave no name and it is confident, the box is counted as a fish on its word alone.
The owner goes through the clips, clicks on each animal and marks it: a fish of one of the five types, a faint one, or something that is not a fish at all. Those marks are the only thing the local model learns from. Each night a new model is trained on every mark so far and set against the one in use, on clips neither has learned from. The new one only takes over if it recognises more of the marked fish with no more wrong calls.
Why spotting fish is hard
- The light keeps changing. Sun and cloud change the whole picture, and patches of light can look like something moving.
- Leaves and debris. A leaf or a piece of pine straw drifting past the lens moves much like a fish, and has been named as one.
- Faint fish. A fish near the edge of what the camera can see is a large dim shape. It is easy to miss, and easy to imagine in empty water.
What you are looking at
- The picture on the front page is replaced every minute.
- The photos, highlights and stats are published about once an hour.
- All times are the pond's own local time.
- The names are not always right. A photo appears here once the owner or Claude's daily review has marked the fish, or two separate answers agreed on it.
- The highlights are clips the owner has starred.