GlueGAN: Gluing Two Images as a Panorama with Adversarial Learning

Abstract

Inspired by panorama photos and Chinese handscroll paintings, we propose an unsupervised algorithm to glue two similar independent images with a generated image as a panorama. Different from existing inpainting methods which can only be used to complete limited missing area with supervised training, the advantage of our GlueGAN lies in generating an image with almost the same size as the input image to form an overall image. To overcome mode collapse which is common in GANs, we propose a two-stage framework. In the first generation stage of HED edge, the model generates simple HED edge images. The second image generation stage aims to generate color image with the aid of HED edge image generated in the first stage. Moreover, we additionally train a doodle-to-edge generator to help users participate in the generation process. To verify our proposed method, a dataset, including 90000 Chinese landscape painting auctions dataset, is created. Experiment results on the dataset show our proposed method can solve this problem well and has greater performance than existing algorithms.