Hand Posture Recognition Using Hidden Conditional Random Fields

Hand posture is one of the natural signs used by people for communication. Thus, there is the need for machines to recognize hand posture. For the recognition of hand posture, two major kinds of model for the object recognition are discussed: bag-of-words model and part-based model. In this thesis, we will review the part-based model in cognition and evaluate a specific computational part-based model proposed by Quattoni et al.: Hidden Conditional Random Fields (HCRFs). Our experiments show that HCRFs are successfully applied on the upright hand posture dataset. In HCRFs, any two nodes are not assumed to be independent and thus may be overlapped. Moreover, global relation of nodes may be incorporated into HCRFs so as to represent large scale dependency among data. Our experiments show that the global feature used by Quattoni et al. is not invariant to in-plane rotation. However, hands are with high degrees of freedom and thus hand postures are frequently in the rotated cases. Therefore, we propose to encode the global relation of nodes by the distance to the image center so as to be invariant to in-plane rotation.