I have trained two eye classifiers (left eye and right eye) with over 6,000 samples. The positive samples are from FERET, BioID, and VALID.
Before I trained my own classifiers. I used the eye classifiers trained by Prof. Modesto F. C. Santana (http://mozart.dis.ulpgc.es/Gias/modesto_eng.html) and Mr. Shameem Hameed (http://umich.edu/~shameem). The classifiers often misclassify eyebrows as eyes, or can not detect eyes when the illumination is not good. To overcome the problems, I trained my own eye classifiers.
Since eyes are not symmetrical as faces, one classifier for each eye is trained. Eyebrows are also included in the target for training. The relationship between pupil locations and eye regions (left and right eye regions) is shown in the following figure. (The left eye is located in top-right region of the face image!)
The classifiers have been tested with our surveillance robot and have satisfying results. They have been contributed to OpenCV. You can download them from: