Learning object model via segment-layout topic

Abstract

Given object instances belonging to the same class, a novel topic model is proposes to learn the part-based object model by a semi-supervised manner. The proposed model, called segment-layout topic model, automatically partitions the instances into several subclasses, discovers the component segments in each instance as the possible parts, and learns the part-based model for each subclass. Unlike the defined parts by high-level object knowledge, the segments focus on the low-level visual features and are easier to be found from the object instances. An iterative process is further proposed to implement model learning. Finally, the proposed method is examined by the experiment of object detection, and is compared with other supervised or semi-supervised methods.