Exploring Contextual Information in a Layered Framework for Group Action Recognition

Contextual information is important for sequence modeling. Hidden Markov models (HMMs) and extensions, which have been widely used for sequence modeling, make simplifying, often unrealistic assumptions on the conditional independence of observations given the class labels, thus cannot accommodate overlapping features or long-term contextual information. In this paper, we introduce a principled layered framework with three implementation methods that take into account contextual information (as available in the whole or part of the sequence). The first two methods are based on state alpha and gamma posteriors (as usually referred to in the HMM formalism). The third method is based on conditional random fields (CRFs), a conditional model that relaxes the independent assumption on the observations required by HMMs for computational tractability. We illustrate our methods with the application of recognizing group actions in meetings. Experiments and comparison with standard HMM baseline showed the validity of the proposed approach.