Yeu-Chern Harn, Vladimir Jojic

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: We present 3C-GAN: a novel multiple generators structures, that contains one conditional generator that generates a semantic part of an image conditional on its input label, and one context generator generates the rest of an image. Compared to original GAN model, this model has multiple generators and gives control over what its generators should generate. Unlike previous multi-generator models use a subsequent generation process, that one layer is generated given the previous layer, our model uses a process of generating different part of the images together. This way the model contains fewer parameters and the generation speed is faster. Speci´Čücally, the model leverages the label information to separate the object from the image correctly. Since the model conditional on the label information does not restrict to generate other parts of an image, we proposed a cost function that encourages the model to generate only the succinct part of an image in terms of label discrimination. We also found an exclusive prior on the mask of the model help separate the object. The experiments on MNIST, SVHN, and CelebA datasets show 3C-GAN can generate different objects with different generators simultaneously, according to the labels given to each generator.