The problem of 3D shape matching is typically restricted to static objects to classify similarity for shape retrieval. In this paper we consider 3D shape matching in temporal sequences where the goal is instead to find similar shapes for a single time-varying object, here the human body. Local-feature distribution descriptors are adopted to provide a rich object description that is invariant to changes in surface topology. Two contributions are made, (i) a comparison of descriptors for shape similarity in temporal sequences of a dynamic free-form object and (ii) a quantitative evaluation based on the Receiver-Operator Characteristic (ROC) curve for the descriptors using a ground-truth data set for synthetic motion sequences. Shape Distribution [25], Spin Image [15], Shape Histogram [1] and Spherical Harmonic [17] descriptors are compared. The highest performance is obtained by volume-sampling shape-histogram descriptors. The descriptors also demonstrate relative insensitivity to parameter setting. The application is demonstrated in captured sequences of 3D human surface motion.
展开▼
