Conditional Image Synthesis with Auxiliary Classifier GANs
Abstract: "In this paper we introduce new methods for the improved training of generative adversarial networks (GANs) for image synthesis. We construct a variant of GANs employing label conditioning that results in 128×128 resolution image samples exhibiting global coherence. We expand on previous work for image quality assessment to provide two new analyses for assessing the discriminability and diversity of samples from class-conditional image synthesis models. These analyses demonstrate that high resolution samples provide class information not present in low resolution samples. Across 1000 ImageNet classes, 128×128 samples are more than twice as discriminable as artificially resized 32×32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data." (Odena et al., 2016)
The authors introduce a new GAN architecture for generating high resolution images from the ImageNet dataset. They show that this architecture makes it possible to split the generation process into many sub-models. They also experimentally demonstrate that generating higher resolution images allow the model to encode more class-specific information, making them more visually discriminable than lower resolution images even after they have been resized to the same resolution.
The second half of the paper introduces metrics for assessing visual discriminability and diversity of synthesized images. The discussion of image diversity in particular is important due to the tendency for GANs to 'collapse' to only produce one image that best fools the discriminator (Goodfellow et al., 2014).
Common methods for image synthesis used today are
The authors propose a conditional GAN that both takes the class to be synthesized as input to the, and includes a classification accuracy term in the loss function of the discriminator. They also split the generation process into many class-specific submodels.
The authors propose two measurement methods to assess the discriminability and diversity of the generated images.
Experimental Results on Image Resolution
Not very different from other GANs. Some unsupported claims about stabilizing training etc.
Discussion of overfitting says b/c nearest neighbours under L1 measure in pixel space are not similar looking it doesn't overfit.
1. Odena, A., Olah, C., & Shlens, J. (2016). Conditional image synthesis with auxiliary classifier gans. arXiv preprint arXiv:1610.09585.