Combining pictorial structure and image features to estimate human pose

This paper estimates human poses in images obtained from daily life or TV. In such conditions, the background, appearance, action, location of humans change variously, therefore, it is challengeable to recognize their poses. Taking advantage of the pictorial structure which describes human as a collection of body parts, this paper turns the problem of human pose estimation to an inference problem in graphic model. Inference messages are described by the information from image and human itself, such as the appearance, segmentation cues and symmetric prior, and are passed by sum-product algorithm. Experiments are made on three public datasets, and considerable accurate results are achieved.