Face Filtering - Insights from Real-World Data

Digital image processing filters continue to be used widely for the normalization of illumination effects in face recognition, both in research and in practice. Their appeal stems from their simplicity, efficiency, predictable and well-understood behaviour, and importantly, lack of catastrophic failure modes. Notwithstanding this widespread use, no work to date has performed a comparative analysis of different filters in challenging, realistic conditions expected in practice - filters in previous work are either adopted in isolation or evaluated in constrained conditions unrepresentative of real-world challenges. In this paper we perform, report, and discuss a comparative evaluation of a number of popular filters on a challenging, real-world data set which contains major changes in illumination, pose (yaw and pitch), camera-user distance, image resolution, and (often neglected) camera type. Our results demonstrate that relative performances of different filters in realistic imaging conditions such as those examined in this paper are vastly different than when the same filters are evaluated in a controlled setting as in previous work. Therefore our results provide important insight for practical application of image filters and future research.