Motion Compression and Retrieval

Human motion capture data are increasingly produced and used in a variety of applications such as 3D animations, digital films, interactive games and medical simulations. The substantial amount of recorded human motion data imposes challenging research problems:

• Because the raw motion capture data consume a significant number of data bits, can we efficiently compress the data without the loss of visual quality?
• Given a large motion data repository, can we efficiently retrieve similar motion sequences based on a motion example?
• And how can we ensure the searched results by computer algorithms are ranked properly?

Assistant Professor Zhigang Deng, graduate students Qin Gu and Qing Li, and a collaborator at Qualcomm, Jingliang Peng, recently presented novel and efficient algorithms for compressing human motion data and retrieving similar motion sequences and ranking the search results in a perceptually consistent order.

The key concept of their algorithms is centered at “motion patterns.” A motion pattern is defined as a representative motion segment for a body part. As such, by representing high-dimensional human motion in a hierarchical scheme, common motion patterns automatically can be extracted from different motion sequences. For example, a walking sequence and a running sequence may share a similar -- or even the same -- upper body movement pattern. By maximally exploiting the local pattern similarity, human motion sequences can be converted to pattern index lists, thus significantly reducing the necessary storage.

Based on the same motion pattern concept, they also developed a perceptually consistent example-based human motion retrieval approach. In this work, the high-dimensional motion retrieval problem is transformed to an integer – in other words, a motion pattern index – matching problem, which can be efficiently solved by extending the classical Knuth-Morris-Pratt (KMP) string match algorithm. Different from the string match case where a limited number of different characters exist, the numbers of motion patterns are typically at thousand levels. Thus, a “quasi-match” concept is introduced into the extended KMP algorithm. The researchers’ extensive comparative user studies showed that this approach is measurably more effective than the state of the art in terms of search accuracy and quality.

For additional information, see papers in Computer Graphics Forum Journal, 28(1), and in the Association for Computing Machinery Special Interest Group on Computer Graphics and Interactive Techniques Symposium on Interactive 3D Graphics and Games 2009.