A comparative study of face re-identification systems under real-world conditions

Face re-identification is largely thought of as a solved problem in the research community, with State-of-the-art systems attaining human-level performance on unconstrained image datasets. However, these results do not seem to translate to the real world. In systems matching people in surveillance-like footage to high-quality images, reported performance is much lower than what the literature would suggest. In this work, we contribute a multi-modal dataset for evaluating real-world performance of facial re-identification systems. We then perform a verification and re-identification evaluation for state-of-art systems on both this dataset and the popular benchmarking dataset Labelled Faces in the Wild (LFW).