Neural rendering in a room

We, as human beings, can understand and picture a familiar scene from arbitrary viewpoints given single image, whereas this is still grand challenge for computers. We hereby present novel solution to mimic such perception capability based on new paradigm of amodal 3D understanding with neural rendering closed scene. Specifically, we first learn the prior knowledge objects in via an offline stage, which facilitates online stage room unseen furniture arrangement. During panoramic image different layouts, utilize holistic neural-rendering-based optimization framework efficiently estimate correct layout deliver realistic free-viewpoint rendering. In order handle domain gap between our method exploits compositional techniques data augmentation training. The experiments both synthetic real datasets demonstrate that two-stage design achieves robust outperforms competing methods by large margin, also show enables various applications, including touring editing. Code are available project webpage: https://zju3dv.github.io/nr_in_a_room/.