Example-Based Retrieval System for Human Motion Data.

This paper proposes a novel motion retrieval system featured with an example-based query interface. The system makes use of a user-created query sequence to search interesting motion in a huge motion database. The target movement of this study is motions of Korean POP (K-POP) dances which are much more complex and dynamic than any other human motions. To cope with those challenging dance motions, we present discriminative pose descriptors based on the overall configuration of a pose and dynamics of interesting joints. As a method to measure motion similarity, subsequence Dynamic Time Warping algorithm is explored that supports partial matches. The proposed intuitive query interface is efficient and especially necessary for the retrieval of K-POP dance motions that are difficult to describe and annotate.