Top down saliency detection via Kullback-Leibler divergence for object recognition

In this paper, we propose a definition of saliency which is useful for object recognition. The desired saliency needs to satisfy the requirements in three levels: 1) representative, it should be a feature of the interest object; 2) discriminative, it needs to be helpful for distinguishing the interest object from the other objects and the background, the usual definitions of saliency for object recognition are able to achieve this goal; 3) easily matched, the saliency of the interest object has to be easy to match correctly. This character, ignored by the usual definitions, is useful for locating the object and estimating the scale and the pose of the object during the recognition. In the proposed definition, the easiness of matching correctly and the discrimination are both measured by the Kullback-Leibler (K-L) divergence. In order to apply the definition to a much broader range of situations, the probability density functions (pdfs) involved in the K-L divergence are not necessary to be restricted to the parametrized families of pdfs, and the K-L divergence is estimated using samples without predicting explicitly the involved pdfs. The experimental results show that the proposed definition of saliency not only is useful for detecting the discriminative features of the interest object, but also improves the accuracy of estimating the support of the object by matching the most salient features.