文摘
In this paper, the problem of unsupervised human action retrieval in 3D mesh sequences is addressed. An action is composed of a mesh sequence, wherein each frame is represented by a static shape descriptor. Six state-of-the-art static descriptors are used to extract meaningful information for each sequence. Firstly, these descriptors are examined in terms of frame-to-frame similarity by means of Receiver Operating Characteristic (ROC) curves. Then, they are utilized in the action retrieval problem, where the query is an entire 3D mesh sequence. Each action is a multidimensional curve which traverses the points defined by the vectors of each descriptor. The estimation of similarity between actions is achieved by calculating the Dynamic Time Warping (DTW) distance between the corresponding curves. The retrieval performance is further examined when the Sakoe band is used to constrain the search space in DTW. The experiments concerning the action retrieval problem were carried out by using a real dataset and an artificial dataset where the proposed retrieval framework is shown to achieve high performance for both datasets.