The task of novel view synthesis refers to generating a new perspective image of the scene through multiple reference images. However, there are occlusions between objects in multi-object scenes, and object information cannot be fully obtained, resulting in artifacts and dislocation problems in the generated new-view scene images. In order to solve this problem, this paper proposed a new perspective image generation network guided by the scene layout map, and annotated a new MONVS. Firstly, it input multiple layout information of the scene and the corresponding camera pose information into the layout prediction module, and calculated the layout information of the scene under a new perspective. Then, it used the bounding box information of the objects marked in the scene to construct an object set of different objects, and used the pixel prediction module to generate the information of each object in the new perspective scene. Finally, it input the obtained new perspective layout and various object information into the scene generator to construct a scene image under the new perspective. Compared with the latest methods on the MONVS and ShapeNet cars data sets, experimental data and visualization results show that in the new perspective image generation of multi-object scenes, the proposed method has good performance on both data sets. It effectively solved the problem of artifacts in the generated image and inaccurate position information of multiple objects in the scene.

[2022-03-30] Accepted Paper /doc/2022.01.0032/v1

[2022-08-05] Printed Article /doc/2022.01.0032/v2