Generalization in holistic versus analytic processing of faces

The distinction between holistic and analytical (or feature-based) approaches to face recognition is widely held to be an important dimension of face recognition research. Holistic techniques analyze the whole face in order to recognize a subject, whereas analytical methodologies are devoted to the processing of different local parts of the face. This paper proposes a principled experimental comparison between these two approaches. Local and global face processing architectures that have access to similar feature representations and classifiers are implemented and tested under the same training and testing conditions. The analysis is performed with a recognition scenario on the difficult BANCA dataset, containing images acquired in degraded and adverse conditions. Different classifiers of increasing complexity are used in each scenario, and different classifier fusion methods are used for combining the local classifers. Our results show that holistic approaches perform accurately only with complex classifiers, whereas feature-based approaches work better with simple classifiers. We were able to show a clear boosting effect by fusing a large number of simple classifiers.