Object categories detection with incorporated geometric context

In this paper, we study object categories detection with a variety of geometric contexts. Usually, object categories are always associated with certain context information. It can help to remove false positive detection. With geometric contextual features, a support vector machine is trained to re-evaluate the initial detection results. Moreover, for the case of that there are determined object categories in an image and the region where an object exists is known, we convert the problem of object categories detection into the one of classification of several object categories. The region can be classified as the one with maximal initial detection score. Alternatively, the detection score for every object category model can be the re-evaluated result of a SVM trained with initial detection score and related geometric context feature. The proposed methods are verified on the dataset of PASCAL VOC 2010. The experimental results demonstrate that accuracy of detection can be improved further with the help of geometric context.