HDP-HCRF for object segmentation

Infinite hidden conditional random fields has been proposed for human behavior analysis which is a non-parametric discriminative model as the extension of HCRF. However, it only model one dimensional temporal relationship by using a chain structure imposed on latent state variables, and would involve huge number of parameters as the number of state increases. In order to solve the 2D object segmentation problem, we propose a novel model relying on hierarchical Dirichlet processes and hidden conditional random fields. Our model maintains properties of non-parametric Bayesian model but with only finite model parameters. Experimental results show the effectiveness of HDP-HCRF on MSRC-21 and VOC 2007 segmentation dataset.