Mining sub-categories for object detection

The visual concept of an object category is usually composed of a set of sub-categories corresponding to different sub-classes, perspectives, spatial configurations and etc. Existing detector training algorithms usually require extensive supervisory information to achieve a satisfactory performance for sub-categorization. In this paper, we propose a detector training algorithm which can automatically mine meaningful sub-categories utilizing only the image contents within the training bounding boxes. The number of sub-categories can also be determined automatically. The mined sub-categories are of medium size and could be further labeled for a variety of applications like sub-category detection, meta-data transferring and etc. Promising detection results are obtained on the challenging PASCAL VOC dataset.