Real-scene-constrained virtual scene layout synthesis for mixed reality

Given a real source scene and a virtual target scene, the real-scene-constrained virtual scene layout synthesis problem is defined as how to re-synthesize the layout of the virtual furniture in the virtual scene to form a new virtual scene such that the new scene not only looks similar to the input real and virtual scenes but also is interactive. The goal of this problem is to maximize interactivity and fidelity which are contradictory. To solve this problem, we propose a real-scene-constrained virtual scene layout synthesis method to synthesize the layout of the virtual furniture in the new virtual scene. We split the scene layout synthesis process into 3 interrelated steps: scene matching, matched furniture layout generating, and unmatched furniture layout generating. For scene matching, we propose a deep scene matching network to predict the matching relationship between real and virtual furniture. For matched furniture layout generating, we propose a layout parameters optimization algorithm to predict suitable layouts of the matched virtual furniture. For unmatched furniture layout generating, we propose a deep scene generating network to predict suitable layouts of unmatched virtual furniture. We evaluate the quality of our method to synthesize scenes of different kinds and sizes. The results show that, compared with the heuristic rules-based method, our method has better matching accuracy and location accuracy. We also design a user study to evaluate the interactivity and fidelity. Compared to the manual method and the heuristic rules-based method, our method has a significant improvement in interactivity and fidelity.