IC-GAN is a set of FAIR models for generating images, objects and scenes on which were absent in the training dataset. IC-GAN can be used for data augmentation.
Generative-adversarial networks (GAN) is a well-proven method for creating both photorealistic and abstract images. However, today these models have an important limitation: they can usually generate images only of those objects or scenes that were originally present in the training dataset. IC-GAN solves this problem by allowing, for example, creating an image with camels surrounded by snow, or with a zebra in an urban environment.
IC-GAN can be used with both annotated and non-annotated datasets. This extends the GAN framework for modeling a mixture of local and overlapping data clusters. IC-GAN can take a single image and then generate images similar to the nearest neighbors of the instance in the dataset. The nearest neighbors are used as input data for the discriminator to force the generator to create samples similar to the samples of the neighborhoods of each input image.
IC-GAN can be used to augment data and include elements or objects that are not usually found in training data. IC-GAN can generate more diverse training data for object recognition models. For example, traditional GAN models will not be able to generate images of zebras in an urban environment, since their training data will most likely contain only images of zebras in the wild.