a geometric approach to landmark detection
This Talk is addressed to the field of computational geometry, from which a fast sorting algorithm with specific mesh structures and a feature finding method of the computer vision is developed. In this Presentation, an approach for an automatic marker detection of a large amount of human body scans will be presented. The data base of the work is composed of incomplete meshes of human body. Depending on the pose of these bodies, we are proposing a method that does not require knowledge about their texture. To do this, the information that will be needed is the shape of these bodies in order to get correspondence between different models. Following the process of this method we will be able to detect the selected markers, which we will call landmarks, to make a registration of a big dataset of scanned people. With a point-to-point correspondence algorithm, it will be possible to deform a similar existing watertight mesh into our scan in order to accurately fill the holes or to compute correspondence for the complete body with the provided landmarks. One of the main obstacles is that we are working with noisy and non-manifold meshes, which are not supported by most geometry libraries. Before using the purpose-built algorithm, we align all data to a fixed position and present and evaluate an algorithm for reducing the noise and aligning all data to a fixed position.