Off the shelf camera based eye pupil center detection has been very popular among computer vision community for the recent years. We propose an accurate and robust regressor based pupil center estimation method without any specialized hardware. The method trains a Support Vector Regressor using HOG features against the Euclidean distance between the center of the train patches and the ground-truth pupil center. On the test stage, we employ a sliding window approach to produce a score image that contains the regressor estimated distances to the pupil center.
We select the best center position among the candidate centers by fitting a second degree polynomial to the maximal score image positions. The detected locations are improved by using an iterative method that repeats the center finding operation until there is no change. We evaluate our method on the challenging BioID data set. The results of the experiments are overall very promising and the system exceeds performance of the similar state of the art methods.