%O Report %F zhang:tech212 %A Zhang, Dongqing %A Chang, Shih-Fu %T Learning Random Attributed Relational Graph for Part-based Object Detection %I Columbia University %X Part-based object detection methods have been shown intuitive and effective in detecting general object classes. However, their practical power is limited due to the need of part-level labels for supervised learning and the low learning speed. In this report, we present a new model called Random Attributed Relational Graph (RARG), by which we show that part matching and model learning can be achieved by combining variational learning methods with the part-based representations. We also discover an important mathematical property relating the object detection likelihood ratio and the partition functions of the Markov Random Field (MRF) in the model. Our approach demonstrates clear benefits over the state of the art in part-based object detection - 2 to 5 times faster in learning with almost the same detection accuracy. The improved learning efficiency allows us to extend the single RARG model to a mixture model for learning and detecting multi-view objects %U http://www.ee.columbia.edu/dvmm/publications/05/techrpt_rarg2005.pdf %8 May %D 2005