FMDistance: A Fast and Effective Distance Function for Motion Capture Data

Kensuke Onuma, Christos Faloutsos, Jessica K. Hodgins

Abstract: Given several motion capture sequences, of similar (but not identical) length, what is a good distance function? We want to find similar sequences, to spot outliers, to create clusters, and to visualize the (large) set of motion capture sequences at our disposal. We propose a set of new features for motion capture sequences. We experiment with numerous variations (112 feature-sets in total, using variations of weights, logarithms, dimensionality reduction), and we show that the appropriate combination leads to near-perfect classification on a database of 226 actions with twelve different categories, and it enables visualization of the whole database as well as outlier detection.