Comparison of face matching techniques under pose variation

The ability to match faces in video is a crucial component for many multimedia applications such as searching and recognizing people in semantic video browsing, surveillance and home video management systems. Unfortunately, most face matching methods were designed for and tested on frontal face images only, which does not comply with the professional and home video scenarios. In video, faces appear at different poses and scales, and the image quality may vary as well. In this paper we analyzed to what extent well-known face matching methods are suitable for matching faces in video. We performed a comparison between the local method Elastic Bunch Graph Matching, the global approaches principle component analysis (PCA) and PCA with linear discriminant analysis (PCA+LDA). The outcome of this study is that while in cases of small face pose variations Elastic Bunch Graph Matching works slightly better, for large face pose variations the global methods provide better performance.