MatchGAN: A Self-supervised Semi-supervised Conditional Generative Adversarial Network

We present a novel self-supervised learning approach for conditional generative adversarial networks (GANs) under semi-supervised setting. Unlike prior approaches which often involve geometric augmentations on the image space such as predicting rotation angles, our pretext task leverages label space. perform augmentation by randomly sampling sensible labels from of few labelled examples available and assigning them target to abundant unlabelled same distribution that ones. The images are then translated grouped into positive negative pairs their labels, acting training involves optimising an auxiliary match loss discriminator's side. tested method two challenging benchmarks, CelebA RaFD, evaluated results using standard metrics including Fr\'{e}chet Inception Distance, Score, Attribute Classification Rate. Extensive empirical evaluation demonstrates effectiveness proposed over competitive baselines existing arts. In particular, surpasses baseline with only 20% used train baseline.