Pal-GAN: Exploring Palette Conditioned Generative Adversarial Networks for Dataset Expansion
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Deep learning is a data-driven field that relies heavily on large quantities of data. GANs are a type of generative model that learn to model data distributions. Conditional variants of GANs exist and conditioning the generator on high information content has been shown to yield high-resolution realistic images. On the other hand, obtaining these “high information” conditions is costly and time-intensive. We approach the problem of image generation using low information conditions and show that adding additional conditions obtained in an unsupervised way in the form of a colour palette can greatly increase the quality of generated images. We also show that expanding datasets with images generated using colour palettes and low information labels does not decrease model performance and may provide a modest increase.