Text-Like motion representation for human motion retrieval

Human motion capture (mo-cap) data has been increasingly applied in animation, movies and games in recent years due to its visual realism, and large amounts of them were accumulated. How to effectively search logically similar motions from large data repositories is a new challenge. The major limitation of existing methods is the semantic level of the features is not high enough to well distinguish different motion categories. In this paper, we propose a text-like motion representation based on key-pose extraction and hierarchical clustering (HC). This motion representation is easy to be extended or combined with topic models to obtain higher semantic-level features for motion retrieval. Our experiments demonstrate its scalability and performance in several applications, including motion retrieval and motion segmentation.