Learning High-Resolution Domain-Specific Representations with a GAN Generator

In recent years generative models of visual data have made a great progress,\nand now they are able to produce images of high quality and diversity. In this\nwork we study representations learnt by a GAN generator. First, we show that\nthese representations can be easily projected onto semantic segmentation map\nusing a lightweight decoder. We find that such semantic projection can be\nlearnt from just a few annotated images. Based on this finding, we propose\nLayerMatch scheme for approximating the representation of a GAN generator that\ncan be used for unsupervised domain-specific pretraining. We consider the\nsemi-supervised learning scenario when a small amount of labeled data is\navailable along with a large unlabeled dataset from the same domain. We find\nthat the use of LayerMatch-pretrained backbone leads to superior accuracy\ncompared to standard supervised pretraining on ImageNet. Moreover, this simple\napproach also outperforms recent semi-supervised semantic segmentation methods\nthat use both labeled and unlabeled data during training. Source code for\nreproducing our experiments will be available at the time of publication.\n