MESSY2NEAT: A Dataset for Image Transfer from Messy Scenes to Neat Scenes by Adversarial Nets

Nowadays, generative adversarial networks (GANs) and its variants are widely used as image-to-image translation tools to generate vivid pictures. However, most of those research efforts focus on changing the images' style, and they are not involved with the location translation. That is, the object locations are kept consistent in original and target images. Accordingly, there is also a lack of datasets suitable for location conversion. In this paper, 1) we present a GANs-based image-to-image translation application involved with location translation. More specifically, we focus on making messy scenes to neat scenes. 2) For the new application scenario, a dataset named MESSY2NEAT is built, which contains six sub-datasets and more than 18 000 paired images. 3) On MESSY2NEAT, three popular GANs-based frameworks, including pix2pix, pix2pixHD, and CycleGAN, are employed to make messy scenes to neat scenes. Experiment results show that it is possible to perform image-to-image translation involved with the location translation using GANs, while more research efforts are needed to improve the results.