Gardias Przemek ; Arthur Eric ; Sun Huaming

Abstract

Although humans perform well at predicting what exists beyond the boundariesof an image, deep models struggle to understand context and extrapolationthrough retained information. This task is known as image outpainting andinvolves generating realistic expansions of an image's boundaries. Currentmodels use generative adversarial networks to generate results which lacklocalized image feature consistency and appear fake. We propose two methods toimprove this issue: the use of a local and global discriminator, and theaddition of residual blocks within the encoding section of the network.Comparisons of our model and the baseline's L1 loss, mean squared error (MSE)loss, and qualitative differences reveal our model is able to naturally extendobject boundaries and produce more internally consistent images compared tocurrent methods but produces lower fidelity images.