A geometric constrained HCRF for object recognition

In the present paper, a discriminative model for object recognition based on the Hidden Conditional Random Fields (HCRF) model is proposed. It impose the constraints on the positions of object parts with a star shape spatial prior. The proposed model can learn the ideal locations of parts, but also their spatial extent. Actually, the added constraints refine the assignment of part labels to local patches. Thus, our model can take advantage of appearance features and geometric structures in recognizing object. Efficient inference and parameter learning approaches are developed to handle the extra hidden variables (i.e. the positions of parts). Experiment results demonstrate the proposed model perform better than original HCRF model.