Hidden Conditional Random Fields for Action Recognition

Abstract

In the field of action recognition, the design of features has been explored extensively, but the choice of action classification methods is limited. Commonly used classification methods like k-Nearest Neighbors and Support Vector Machines assume conditional independency between features. In contrast, Hidden Conditional Random Fields (HCRFs) include the spatial or temporal dependencies of features to be better suited for rich, overlapping features. In this paper, we investigate the performance of HCRF and Max-Margin HCRF and their baseline versions, the root model and Multi-class SVM, respectively, for action recognition on the Weizmann dataset. We introduce the Part Labels method, which uses explicitly the part labels learned by HCRF as a new set of local features. We show that only modelling spatial structures in 2D space is not sufficient to justify the additional complexity of HCRF, MMHCRF or the Part Labels method for action recognition.