An experimental study on content-based face annotation of photos

Abstract

Face annotation of photos, a key enabling technology for many exciting new applications, has been gaining broad interest. The task is different from the general face recognition problem because the dataset is not constrained — an unlabelled face may not have any corresponding match in the training set. Moreover, faces in real-life photos have a significantly wider variation range than those in the conventional face datasets. We designed and conducted a thorough experimental study to understand the efficacy of face recognition methods for annotating faces in real-world scenarios. The findings of this study should provide information for various design choices for a practical and high-accuracy face annotation system.