Comparison of Skeleton and Non-Skeleton Shape Descriptors for 3D Video

This paper presents a performance evaluation of shape similarity metrics for 3D video sequences of people with unknown temporal correspondence. Previous evaluation focuses on non-skeletal similarity metrics. Since the human motion is essentially an articulated motion, it is also interesting to investigate skeletal-based similarity metrics. In this paper, we extend previous evaluation of non-skeletal similarity metrics to skeletal-based similarity metrics. A quantitative evaluation based on the Receiver-Operator Characteristic (ROC) curve for the descriptors using a ground-truth dataset for synthetic motion sequences is presented. Reeb Graph, Multi-Dimensional Scaling and Shape Histograms are compared with and without a temporal filter. Reeb Graph and Shape Histograms achieve comparable ROC performance, both outperform MDS in the task of finding similar poses of the same person in 3D video. Finally, temporal Reeb Graph and Shape Histograms are applied to a public database of 3D video of people to identify optimal transitions and synthesise 3D character animation. Results demonstrate the accurate matching of surface shape and motion.