Recognition and visual learning of articulated shape by accumulative Hopfield matching

We describe a system that can recognize and learn a visual model of an articulated object automatically given different views of the object, provided that the local structure is unchanged. The system is based on the Hopfield style network and finds the feature correspondences between different views of an articulated object. With this proposed matching system, we can finally learn the relationship between articulated parts of the object and the poses detected. Experiments on real images show the effectiveness of the proposed system