Is 2D Unlabeled Data Adequate for Recognizing Facial Expressions?

Abstract

Automatic facial expression recognition is one of the important challenges for computer vision and machine learning. Despite the fact that many successes have been achieved in the recent years, several important but unresolved problems still remain. This paper describes a facial expression recognition system based on the random forest technique. Contrary to the many previous methods, the proposed system uses only very simple landmark features, with the view of a possible real-time implementation on low-cost portable devices. Both supervised and unsupervised variants of the method are presented. However, the main objective of the paper is to provide some quantitative experimental evidence behind more fundamental questions in facial articulation analysis, namely the relative significance of 3D information as oppose to 2D data only and importance of the labelled training data in the supervised learning as opposed to the unsupervised learning. The comprehensive experiments are performed on the BU-3DFE facial expression database. These experiments not only show the effectiveness of the described methods but also demonstrate that the common assumptions about facial expression recognition are debatable.