Recognition of Visual Dynamical Processes: Theory, Kernels, and Experimental Evaluation
MetadataShow full item record
Over the past few years, several papers have used Linear Dynamical Systems (LDS)s for modeling, registration, segmentation, and recognition of visual dynamical processes, such as human gaits, dynamic textures and lip articulations. The recognition framework involves identifying the parameters of the LDSs from features extracted from a training set of videos, using metrics on the space of dynamical systems to compare them, and combining these metrics with different classification methods. Usually, each paper makes an ad-hoc choice for every step, and tests the recognition framework on small data sets often involving only one application. We present a detailed evaluation of the LDS-based recognition pipeline; comparing identification methods, metrics, and classification techniques. We propose new metrics that have certain invariance properties and explore a number of variations to the existing metrics. We perform experimental evaluations on well-known data sets of human gaits, dynamic textures, and lip articulations and provide benchmark recognition results. We also analyze the robustness of the recognition pipeline with respect to changes in observation and experimental conditions. Overall, this work represents the most extensive to-date evaluation of the LDS-based recognition framework.