Introducing Statistical Shape Prior to Probabilistic Contour Extraction

In this paper, statistical shape prior and smoothness of contours are simultaneously incorporated into probabilistic contour estimation. By means of principal component analysis, shape samples are generated from sampling a density in a lower dimensional space. Based on the state samples derived from the shape samples and on the prior density encoding smoothness, the contour with a specific shape is estimated by a sequential Monte-Carlo method. Extensive experiments on the real-world hand images show that the proposed approach significantly improves the robustness to cluttered background, occlusion and noise with respect to the objects with specific shapes.