MT-GAN: toward realistic image composition based on spatial features

Abstract

The purpose of image composition is to combine the visual elements of different natural images to produce a natural image. The performance of most existing image composition methods drops significantly when they solve multiple issues, such as image harmonization, image blending, shadow generation, object placement, and spatial transformation. To address this problem, we propose a multitask GAN for image compositing based on spatial features, aiming to simultaneously address the geometric and appearance inconsistency. We use three related learning objective functions to train the network. Moreover, a new dataset including 7756 images with RoI region annotations is contributed to help evaluate the multitask image compositing results. Extensive experiments demonstrate that our proposed method is effective on our dataset.