2019/10/16 by Nikolay Jetchev, Urs Bergmann, Jetchev, Nikolay +3 · 2 citations
Computer Science · #Generative Adversarial Networks and Image Synthesis #Advanced Vision and Imaging #Computer Graphics and Visualization Techniques
paper · pdf · doi:10.48550/arxiv.1910.07236
Cutting and pasting image segments feels intuitive: the choice of source\ntemplates gives artists flexibility in recombining existing source material.\nFormally, this process takes an image set as input and outputs a collage of the\nset elements. Such selection from sets of source templates does not fit easily\nin classical convolutional neural models requiring inputs of fixed size.\nInspired by advances in attention and set-input machine learning, we present a\nnovel architecture that can generate in one forward pass image collages of\nsource templates using set-structured representations. This paper has the\nfollowing contributions: (i) a novel framework for image generation called\nMemory Attentive Generation of Image Collages (MAGIC) which gives artists new\nways to create digital collages; (ii) from the machine-learning perspective, we\nshow a novel Generative Adversarial Networks (GAN) architecture that uses\nSet-Transformer layers and set-pooling to blend sets of random image samples -\na hybrid non-parametric approach.\n