Pixel-Level Domain Transfer
We present an image-conditional image generation model. The model transfers an input domain to a target domain in semantic level, and generates the target image in pixel level. To generate realistic target images, we employ the real/fake-discriminator as
- PDF / 2,903,783 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 86 Downloads / 168 Views
KAIST, Daejeon, South Korea {dgyoo,nikim}@rcv.kaist.ac.kr, {sunggyun,iskweon}@kaist.ac.kr 2 Lunit Inc., Seoul, South Korea [email protected]
Abstract. We present an image-conditional image generation model. The model transfers an input domain to a target domain in semantic level, and generates the target image in pixel level. To generate realistic target images, we employ the real/fake-discriminator as in Generative Adversarial Nets [6], but also introduce a novel domain-discriminator to make the generated image relevant to the input image. We verify our model through a challenging task of generating a piece of clothing from an input image of a dressed person. We present a high quality clothing dataset containing the two domains, and succeed in demonstrating decent results.
Keywords: Domain transfer
1
· Generative Adversarial Nets
Introduction
Every morning, we agonize in front of the closet over what to wear, how to dress up, and imagine ourselves with different clothes on. To generate mental images [4] of ourselves wearing clothes on a hanger is an effortless work for our brain. In our daily lives, we ceaselessly perceive visual scene or objects, and often transfer them to different forms by the mental imagery. Our focus of this paper lies on the problem; to enable a machine to transfer a visual input into different forms and to visualize the various forms by generating a pixel-level image. Image generation has been attempted by a long line of works [9,21,24] but generating realistic images has been challenging since an image itself is high dimensional and has complex relations between pixels. However, several recent works have succeeded in generating realistic images [6,8,22,23], with the drastic advances of deep learning. Although these works are similar to ours in terms of image generation, ours is distinct in terms of image-conditioned image generation. We take an image as a conditioned input lying in a domain, and re-draw a target image lying on another. Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46484-8 31) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VIII, LNCS 9912, pp. 517–532, 2016. DOI: 10.1007/978-3-319-46484-8 31
518
D. Yoo et al.
A source image.
Possible target images.
Fig. 1. A real example showing non-deterministic property of target image in the pixellevel domain transfer problem.
In this work, we define two domains; a source domain and a target domain. The two domains are connected by a semantic meaning. For instance, if we define an image of a dressed person as a source domain, a piece of the person’s clothing is defined as the target domain. Transferring an image domain into a different image domain has been proposed in computer vision [1,7,10,12,16,20], but all these adaptations take place in the feature space, i.e. the model parameters are adapted. However, our method directly produces target images. We transfer a knowledge in a
Data Loading...
