Toward More Realistic Face Recognition Evaluation Protocols for the YouTube Faces Database

One of the key factors to measure the progress of a research problem is the design of appropriate evaluation protocols defined on suitable databases. Recently, the introduction of comprehensive databases and benchmarks of face videos has had a great impact on the development of new face recognition techniques. However, most of the protocols provided for these datasets are limited and do not capture requirements of unconstrained scenarios. That is why sometimes the performance of face recognition methods on current benchmarks seems to be saturated. To address this lack, the tendency is to collect new datasets, which is more expensive and sometimes the main the problem is not the data but the protocols. In this work, we propose new relevant evaluation protocols for the YouTube Faces database (REP-YTF) supporting face verification and open/closed-set identification. The proposal better fits realistic face recognition scenarios and allows us to test existing algorithms at relevant assessment points, under different openness values and taking both videos and images as the gallery. We provide an extensive experimental evaluation, by combining several well-established feature representations with three different metric learning algorithms. The obtained results show that by using the proposed evaluation protocols, there is room for improvement in the recognition performance on the YouTube Faces database.