Max Margin AND / OR Graph Learning for Efficient Articulated Object

In this paper we formulate a novel AND/OR graph representation for parsing articulated objects into parts and recovering their poses. The AND/OR graph allows us to handle an enormous variety of articulated poses with a compact graphical model. We develop a novel inference algorithm, compositional inference, that uses a bottom-up compositional process for proposing configurations for the object. The strategy of surround suppression is applied to ensure that the inference time is polynomial in the size of input data. We present a novel structure-learning method, Max Margin AND/OR Graph (MM-AOG), to learn the parameters of the AND/OR graph model discriminatively. Maxmargin learning is a generalization of the training algorithm for support vector machines (SVMs). The parameters are optimized globally, i.e. the weights of the appearance model for individual nodes and the relative importance of spatial relationships between nodes are learnt simultaneously. The kernel trick can be used to handle high dimensional features and to enable complex similarity measures to discriminate between object configurations. We applied our approach – Long (Leo) Zhu Department of Statistics University of California at Los Angeles Los Angeles, CA 90095 E-mail: [email protected] Yuanhao Chen University of Science and Technology of China Hefei, Anhui 230026 P.R.China E-mail: [email protected] Chenxi Lin Microsoft Research Asia E-mail: [email protected] Alan Yuille Department of Statistics,Psychology and Computer Science University of California at Los Angeles Los Angeles, CA 90095 E-mail: [email protected] the AND/OR graph representation, compositional inference and max-margin learning – to the tasks of detecting, segmenting and parsing horses and human body. We demonstrate that the inference algorithm is fast and analyze its computational complexity empirically. To evaluate max margin learning, we perform comparison experiments on the horse and human baseball datasets, showing significant improvements over state of the art methods on benchmarked datasets.