OpenGAN: Open Set Generative Adversarial Networks

Many existing conditional Generative Adversarial Networks (cGANs) are limited to conditioning on pre-defined and fixed class-level semantic labels or attributes. We propose an open set GAN architecture (OpenGAN) that is conditioned per-input sample with a feature embedding drawn from metric space. Using state-of-the-art learning model encodes both fine-grained information, we able generate samples semantically similar given source image. The information extracted by the transfers out-of-distribution novel classes, allowing generative produce outside of training distribution. show our proposed method 256\(\times \)256 resolution images classes visual quality those classes. In lieu image, demonstrate random sampling space also results in high-quality samples. interpolation latent visually plausible transformations image Finally, usefulness generated downstream task data augmentation demonstrated. classifier performance can be significantly improved augmenting OpenGAN