The result is http://eyemap.io. It's a simple-to-use, affordable gaze tracking analytics system with your tablet. I'm still testing if people actually want the service. It's hard to make sense of it, because people sign up to the mailing list but nobody would pay for it.
How accurate is the tracking? Recently tried to do gaze tracking on an iPad for a hackathon and whilst we were able to detect viewing direction (left, right, up, down) getting an actual pixel info was nearly impossible. We had stable and good eye corner and pupil detection, the problem was the movements were just so small the data got extremely noisy.
I used something like 4 different algorithms in the end to approximate the pixel location. A voting algorithm + smoothing algorithm is also used to find the final pixel location.
What I did to measure accuracy was to draw targets on the screen, and focus my eyes on them. The error the absolute distance from the estimated pixel to the actual target pixel.
To answer your question - on a Nexus 10, there is a expected error of about 0 to 160 pixels radius in office lighting conditions.
As for small movements, I use the full resolution video of 1344xwhatever, which gives a lot more leeway in terms of movements. The movements are then smoothed over time using a moving average over 8 frames, and another smoothing algo uses a pyramid kernel.
TL;DR: lots of algorithm. Quite a number are dodgy.
Can you make an online demo or something that shows it working on my laptop? I have a hard time believing that you can localize my gazes - I need to see for myself.
Unless you are seeing keen interest from people willing to pay for this directly in its current form, where I would research and learn how website or app optimization platforms market themselves.
This is mine. 2 weeks ago I set out to learn how to write android apps, and do gaze tracking. Story here: http://blog.chewxy.com/2014/04/08/eyetracking-jetpack-joyrid...
The result is http://eyemap.io. It's a simple-to-use, affordable gaze tracking analytics system with your tablet. I'm still testing if people actually want the service. It's hard to make sense of it, because people sign up to the mailing list but nobody would pay for it.