Neural Scene Decoration from a Single Photograph

AbstractFurnishing and rendering indoor scenes has been a long-standing task for interior design, where artists create conceptual design the space, build 3D model of decorate, then perform rendering. Although is important, it tedious requires tremendous effort. In this paper, we introduce new problem domain-specific scene image synthesis, namely neural decoration. Given photograph an empty space list decorations with layout determined by user, aim to synthesize same desired furnishing decorations. Neural decoration can be applied designs in simple yet effective manner. Our attempt research novel generation architecture that transforms object into realistic furnished photograph. We demonstrate performance our proposed method comparing conditional synthesis baselines built upon prevailing translation approaches both qualitatively quantitatively. conduct extensive experiments further validate plausibility aesthetics generated scenes. implementation available at https://github.com/hkust-vgd/neural_scene_decoration.KeywordsGANsImage synthesisIndoor