UHDB31: A Dataset for Better Understanding Face Recognition Across Pose and Illumination Variation

Ha A. Le, Ioannis A. Kakadiaris; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2555-2563

The face recognition accuracy achieved on current benchmark datasets is saturated. Although multiple face datasets have been published recently, they only focus on the number of samples and lack diversity on facial appearance factors, such as pose and illumination. In addition, while 3D data have been demonstrated improved face recognition accuracy by a significant margin, only a few 3D face datasets provide high quality 2D and 3D data. In this paper, we introduce a new and challenging dataset, called \dbname, which not only allows direct measurement of the influence of pose, illumination, and resolution on face recognition but also facilitates different experimental configurations with both 2D and 3D data. We conduct a series of experiments with various face recognition algorithms and point out how far they are from solving the face recognition problem under pose, illumination, and resolution variation. The dataset is publicly available and free for research use.