Intelligent synthesis driven model calibration: framework and face recognition application

Deep Neural Networks (DNNs) that achieve state-of-the-art results are still prone to suffer performance degradation when deployed in many real-world scenarios due to shifts between the training and deployment domains. Limited data from a given setting can be enriched through synthesis, then used to calibrate a pre-trained DNN to improve the performance in the setting. Most enrichment approaches try to generate as much data as possible; however, this blind approach is computationally expensive and can lead to generating redundant data. Contrary to this, we develop synthesis, here exemplified for faces, methods and propose information-driven approaches to exploit and optimally select face synthesis types both at training and testing. We show that our approaches, without re-designing a new DNN, lead to more efficient training and improved performance. We demonstrate the effectiveness of our approaches by calibrating a state-of-the-art DNN to two challenging face recognition datasets.