Contour-HOG: A Stub Feature based Level Set Method for Learning Object Contour

An object can be effectively characterized by its contour. Caselles et al. [1] introduced the concept of geodesic active contours, which applies the energy reducing form to acquire contours. Shape priors are great helpful to obtaining more accurate contours. Leventon [6] utilized the curvature prior as the shape prior for different classes of objects to guide contour evolution. Etyngier et al. [3] proposed a non-linear manifold learning method for learning shape prior. Another line of work uses edges to describe objects which provide local perspective of an object and are robust when part of the object is occluded. However, since the global perspective is missing, the arrangement of the edge features, such as the pairwise interactions between edge features [2, 5] or the relative positions of edge features with respect to the centroid of the shape [7], is exploited to improve the edge based models. We propose a novel edge-based method for learning objects. Given an image, our method first detects edgelet feature as a rough contour for an object. Edgelet feature indicates potential positions for the contour and may stop curve evolution. These positions are referred to as stub features. Object contour is adaptively refined by the level set method. The evaluation criteria for contour evolution is defined by the similarity between the evolving contour and the target contour computed by their HOG features. Therefore the curve evolution method is referred to as the Contour-HOG method. We formulate the joint distribution of the edgelet feature, the HOG feature and the curvature of the evolved contour in a probabilistic model, and perform classification by computing the posterior of the evolved contour conditioned on the three types of features. Compared with previous methods, our method uses stub features to roughly localize a target object. This allows us to accurately capture the contour of the object. Moreover, the method fuses both local and global features to better describe the contour and thus improves the recognition accuracy. Our method begins by detecting edgelet feature [8]. We use this feature to roughly find an object in an image. With the detected stub feature, we compute their similarities with the stub feature in training data under a predefined edge mask Mk,s. The similarity is computed as