Sahneyi koruyan kişi görünüm aktarımı

Recent developments in deep learning and deep generative models have enabled numerous new applications in the area of image and video editing. One such emerging topic is the editing of pose and appearance of the person across images. Several recent works have introduced methods for transforming the pose of a single person within the same scene. In addition, there are works that aim to model the appearance, scene and pose through separate representations, and produce new images with arbitrary appearances, backgrounds or poses through these representations. Similarly, some works model the appearance of a particular person through his/her video to learn to generate the person in arbitrary poses. In this thesis, we tackle a comprehensive version of these problems by aiming to learn a generative model that can replace the person in an image with another person from a single image, both with possibly different poses and different backgrounds. More specifically, we aim to learn a generative model that takes a single image of an actor with an arbitrary pose and a stuntman that provides the pose and scene information and yield a new image that contains the background scene and pose of the stuntman and the appearance of the actor. We aim to obtain a realistic final image, which requires properly re-generating the actor appearance in the pose of the stuntman and synthesizing the missing background and foreground pixel information due to pose and physical characteristic differences. For this purpose, we propose an end-to-end framework that is trained to maximize the prediction quality of pixel-wise foreground and background details via masked reconstruction loss terms and realism of the output image via an adversarial trained discriminator network. We also introduce a new benchmark by adapting the video segmentation datasets YouTube VOS and Davis for the proposed task. We experimentally investigate and evaluate our approach on the proposed benchmark dataset.