Image Synthesis From Hand Drawn Reconfigurable Layouts

Abstract

Image synthesis is a significant computer vision problem with numerous applications. With the rise of Generative Adversarial Networks, there has been a significant advancement in this area (GANs). Recent times have seen a rise in interest towards conditional image generation from layout. To create useful applications with a userfriendly interface, taking control of the image generating process is essential. The focus is to study generative models for generating almost real images from the spatial layout in which bounding boxes of objects and their categories are configured in an image lattice, and style codes (i.e., latent vector encoding structural variation). The study of intuitive paradigm for the problem, layout to mask to image is done. TO connect the dap between input layout and synthesized images, layout to mask component major role as it deeply interacts with the generator network. A GAN is built for layout to mask to image synthesis with style control and layout control at both object level and image level. The controllablility is realised by ISLA Norm (Instance Sensitive and Layout Aware Normalization) scheme. We create and experiment on a the challenging Visual Genome dataset. 1