Precise Eye Location Using Dempster-Shafer Theory

Eye location is an important step in automatic visual interpretation and face recognition. In this paper, we present a novel eye location algorithm based on Dempster-Shafer's evidential reasoning. Four eye detectors are trained by AdaBoost with different combinations of feature spaces and samples. They detect face region respectively and produce an eye candidate set, then two of the four eye detectors calculate the confidence of every eye candidate. The confidence is converted to belief and plausibility. We use the combining rule to obtain combined belief and plausibility of each eye candidate, which represent the fusion information of the two eye detectors. The centers of eye candidates with great plausibility are considered as actual eye centers. Experimental results on several open face databases demonstrate that our method is precise, robust and has less computational complexity than other newly proposed ones.