Layout Structure Assisted Indoor Image Generation

Abstract

The existing methods can generate images in accord with scene graph, but the obtained images may appear blurs at the edges and disorders in the structure, due to the lacks of the structure information. In this paper, by considering the indoor images contain more layout structures than outdoor ones, we focus on the indoor image generation assisted with the layout structures. In the proposed method, through fusing the scene graph features together with the layout structure, the graph convolutional network is employed to convert the fused semantic information into the feature representation of scenes. Subsequently, a refined encoder-decoder network is also used for generating the final images. In the experiments, we compare the proposed method with the existing works on the indoor image dataset in terms of subjective and objective evaluations. The experimental results show that our method can achieve better IoU metric, and the visualized results also illustrate that the proposed approach can generate more clear indoor images with better layout structures.