Probabilistic Graphical Models for Object Recognition

Kurzfassung

Object category recognition is one of the most challenging topics in today’s computer vision. One approach which shows promising results is the ensemble of classifiers. The main benefits of such methods are the combination of task-specific expertise as well as the ability to combine multi-modal information. In this work, a new ensemble method for the task of category recognition in different environments is presented. The focus is on service robotic perception in an open environment, where the robot’s task is to recognize objects of predefined categories, based on an a priori training on a public database. The proposed concept of classifier fusion is based on the theory of a Markov Random Field (MRF). Hereby, the undirected graphical model consists of vertices representing the experts’ class-predictions and the final prediction. Since the number of the modeled relations between the nodes has a large effect on recognition accuracy and computational performance, different architectural structures are presented. By exploiting the specific characteristics of a MRF, a hybrid expert fusion method in terms of inference techniques is developed. Besides the common inference, which is considering all available experts, the MRF ensemble method can also be executed as a Dynamic Classifier Selection (DCS) system. In the experiments, the committee-dependent performance boost/decline of the proposed method, as well as the effect of a multi-modal ensemble are shown. On the basis of these observations, the benefits of the efficient dynamic inference technique are illustrated. Besides the reduction of computational effort, the ensemble either performs on the same or, in case of a strong committee, above the level of the best single expert. Finally, the impact of a cross domain application is analyzed which leads to general recommendations for adaptation in ensemble methods. Large scale evaluations were performed on two publicly available RGB-D datasets.