Nakazawa et al. introduce an exciting new technique that takes advantage of an aspherical model of the eye's corneal surface to register an image of a scene reflected off of the cornea (an "eye reflection" image) with an image of the actual scene itself. Despite many efforts, the registration of these types of images continues to be difficult due to challenges in dewarping nonlinear distortions in the reflected scene image and the presence of "noise" (e.g., iris features, eyelid and eyelash shadows) in reflected images. The authors describe a new 2-step registration approach that builds on previous feature-point-based alignment techniques that can be sensitive to image noise or require uniform threshold values to assess how well a transformation algorithm dewarps the reflected image. First, a coarse registration is performed using a single-point registration and random resample consensus algorithm, enabling the technique to work well even when applied to images containing large amounts of noise. A fine registration step is then performed by comparing many pairs of points in the reflected and actual scene images to determine a final dewarping function and increase the accuracy of the technique. The algorithm was experimentally verified in 4 subjects after simultaneously collecting reflected and actual scene images from 3 indoor and 2 outdoor scenes and achieved a mean accuracy of approximately 1º of error in registration between reflected and actual scene images. While this value is slightly greater than typical accuracies of current video eye trackers and this technique assumes that the eye’s optical axis corresponds to the direction of gaze (or line of sight), the newly introduced method is exciting as it improves upon current eye gaze trackers by potentially allowing natural scene illumination to replace the infrared LEDs typically used in video eye trackers to illuminate the cornea and providing a parallax-free technique that requires no calibration. In addition to applications that estimate point of gaze, this technique could prove valuable for improving the performance and robustness of iris recognition and bioinformatics techniques by providing the ability to remove a reflected scene image from an iris image to better discern actual iris features.
You must log in
to add comments.