Dataset Modification Captured in Uncontrolled Conditions

There are many computer vision tasks where even the state-of-the-art object detection methods are not able to yield optimal results. The most interesting element of today are Generative Adversarial Networks. These networks are able to handle various challenging transformation or difficult generative tasks. In this paper, we will try to improve the success of state of the art of detection models by applying appropriate segmentation and transformation on historical images. We have modified a set of historical images using GAN networks. Then we have compared object detectors on the original images with the new generated ones. The main motivation of the research is to improve the facial biometric system, which highly depends on detection phase. At the end we have compared the outputs from non-face detectors before and after dataset modification.