Context-based motion retrieval using vector space model

Abstract

Motion retrieval is the problem of retrieving highly relevant motions in a timely manner. The principal challenge is to characterize the similarity between two motions effectively, which is tightly related to the gap between the motion data's representation and its semantics. Our approach uses vector space model to measure the similarities among motions, which are made discrete using the vocabulary technique and transformation invariant using the relational feature model. In our approach, relational features are first extracted from motion data. then such features are clustered into a motion vocabulary. Finally motions are turned into bag of words and retrieved using vector-space model. We implemented this new system and tested it on two benchmark databases composed of real world data. Two existing methods, the dynamics time warping method and the binary feature method, are implemented for comparison. The results shows that our system are comparable in effectiveness with the dynamic time warping system, but runs 100 to 400 times faster. In comparison to retrieval with binary features, it is just as fast but more accurate and practical.The success of our system points to several additional improvements. Our experiments reveal that the velocity features improve the relevance of retrieved results, but more effort should be dedicated to determining the best set of features for motion retrieval. The same experiments should be performed on large databases and in particular to test how this performance generalizes on test motions outside the original database. The alternative vocabulary organizations, such as vocabulary tree and random forest, should be investigated because they can improve our approach by providing more flexibility to the similarity scoring model and reducing the approximation error of the vocabulary. Because the bag of words model ignores the temporal ordering of key features, a wavelet model should also be explored as a mechanism to encode features across different time scales.

(cont.) The alternative vocabulary organizations, such as vocabulary tree and random forest, should be investigated because they can improve our approach by providing more flexibility to the similarity scoring model and reducing the approximation error of the vocabulary. Because the bag of words model ignores the temporal ordering of key features, a wavelet model should also be explored as a mechanism to encode features across different time scales.