Object localization based on discriminative visual words

This paper aims at learning discriminative visual words for object localization. These visual words are different from those learned from the generic object recognition which usually contain the negative visual words located on the background. For the purpose of object localization, the approach requires that the positive discriminative visual words mostly lie in the foreground object and the negative ones lie in the background. We firstly rank the visual words by the following three methods: the SVM classifier, the foreground likelihood ratio and the mutual information. Then, we integrate the three ranking results in an optimal combinational way and select the discriminative visual words by maximizing the object hit rate. Moreover, a coarse-to-fine detection framework to locate the object is designed. In the first stage, a branch-and-bound scheme combined with the discriminative visual words is implemented to find candidate object regions. In the second stage, a sliding window classifier is used to find the object location. The experimental results demonstrate that the approach is effective and efficient, and superior to Efficient Subwindow Search scheme.