Abstract
In this paper, the geometry-aware GAN, referred to as GAGAN, is proposed to address the issue of face attribute transfer with unpaired data. The source and target images are not aligned and come from different individuals. The key idea is to deform the source image according to the geometry features and generate a high-resolution image with the desired attribute. To address the problem of the unpaired training samples, the CycleGAN architecture is applied to form an attribute adding and removing cycle, where the bilateral mappings between the source and target domains are learned. The geometry flow and occlusion mask are learned by the warping sub-network to capture the geometric variation between the two domains. In the attribute adding process, the spatial transformer network (STN) warps the source face into the desired pose and shape according to the flow, and the transfer sub-network hallucinates new components on the warped image. In the attribute removing process, the recover sub-network and the STN reverts the sample back to the source domain. Experiments on the benchmarks CELEBA and CELEBA-HQ datasets demonstrate the advantages of our method compared to the baselines, in terms of both quantitative and qualitative evaluation.
Highlights
Understanding and manipulating face images has drawn much attention in the field of computer vision and graphics
The first category of works mainly rely on computer graphics techniques, which obtain the manipulated image by finding the geometric warping between the source and target images, or they reuse sample patches of existing images
We propose an end-to-end geometry-aware generative adversarial networks (GANs) (GAGAN), which learns a dual generator to perform unpaired image-to-image translation for different domains
Summary
Understanding and manipulating face images has drawn much attention in the field of computer vision and graphics. A large amount of the literature has explored the techniques of face frontalization [1]–[3], face superresolution [4], face aging [5], face animation [6]–[8], and etc. These methods fall in two schools of methods. The second category of works apply computer vision techniques, such as variational autoencoders (VAEs) [9] and generative adversarial networks (GANs) [10]. Reconstructing the high-resolution face with fewer artifacts while preserving the identity remains a challenge
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.