One set of frames, many ways of reading them. Every tracker below is taught from the same looks and tested on the same other looks, so the table shows which method reads the eyes best on this camera, in this light, for this person.
1Where the frames come from
A simulated person needs no camera: an eyeball that turns to look at a point, eyelids, blinks, a head that drifts, and a face model that is a little wrong, as the real one is. Move the mouse and it looks there. It is for seeing how the methods behave, not for real numbers: those come from the camera.
2What the face model is given
eye corners and lids the face model's iris points after a test: the 3D eyeballs and gaze lines of JEOresearch's tracker
Convert the picture first
3The eye, converted
Each eye is cut out of the camera picture, turned level, made the same size and put in black and white. Everything after that works on this small standard picture.
the face model's iris point our iris finder
4The trackers, compared
The test teaches, then scores. To teach, either a chick appears at one place after another (look at it until it moves), or a bee flies slowly to a flower (follow the bee with your eyes; a flight your eyes did not follow is left out). Then nine chicks, at places no tracker has seen, score them all. Then Try it full screen: look at a tile and hold, or drag a bee onto a flower with your eyes. A recording holds the small eye pictures and face points of whoever sat the test, never the whole camera picture; it stays on this device unless you save and share the file.
Tests this visit
Run once with each way of teaching, in the same light and seat, to see which teaches better for this person. Simulated runs are marked.
Where each tracker thought the test looks were
Crosses are where the chick was. A dot is where a tracker placed that look; the line is its error. Tick a tracker in the table to show it.
Live
One dot for each ticked tracker, where it thinks the eyes are now.
5A face movement as the switch
For a child who cannot aim with the eyes or the head, a movement of the face can be the switch: the board lights each tile in turn and the movement chooses. No calibration. What matters is not whether the camera sees the movement but whether the child can make it when they mean to, and only then. This test asks for each movement a few times, and watches for it while the face is meant to be still.