I think this is impressive work. My lab does eye tracking with really expensive equipment locally. It would be interesting to conduct visual/face processing studies through the web and include eye tracking data with this library. Granted the accuracy is low compared to specialized equipment and predictions are highly variable, other contributors could enhance its accuracy. I suspect that the variability in shape, size, and structure of human faces contributed greatly to the high variance/low resolution predictions issue. Maybe categorizing the user's face shape a priori would enhance predictions. That is, the library would first estimate the "type of face" and use those data to inform the eye tracking. This might not make sense--just a thought.