Title: The use of landmarks within elastic planar shape analysis
Authors: Justin Strait - University of Georgia (United States) [presenting]
Abstract: In shape analysis, landmarks are generally identified as important features of the shape which are mathematically and/or anatomically relevant. While recent advances in the field have been focused on shape representations which treat the underlying contour as an object in infinite-dimensional space, we will discuss some uses of landmarks as possibly useful inferential tools in this context. In the unsupervised learning scenario, landmarks can be thought of as a low-dimensional set of points along shape contours which ``approximate'' the underlying object well. While other methods may produce estimates of underlying features, the advantage of the model-based method we propose is the ease of interpretability in identifying landmarks as these important features. If class labels are also known in addition, we can treat landmarks as latent variables, in the sense that we can identify them as important points which discriminate between shape classes optimally. In both settings, we propose a hierarchical, model-based approach, producing estimates of uncertainty for landmark locations. We will discuss modeling issues which arise from the Bayesian perspective, and demonstrate the use of these models on both simulated and real data.