Scene Editing Based on NeRF

The NeRF has achieved impressive results in novel view synthesis and 3D reconstruction, but beyond recovering the geometric structure of the scene, there is a need for more tasks to scene understanding and interaction. Therefore, in this paper, we propose a new scene editing technique by generalizing pixel semantics and colors rendering formulas, which can achieve the unique displays of the specific semantic targets or masking them. So far, most NeRF models have been designed to learn the entire scene. However, When there are many objects in the scene and the background is complex, it often leads to longer learning time, poorer rendering performance, and even many artifacts. Therefore, using the proposed scene editing technique, this article focuses NeRF on learning specific objectives without being affected by complex backgrounds. It results in faster training speed and greater rendering quality. Finally, to address the problem of incorrect inference in unsupervised regions of the scene, we design a self-supervised loop combining morphological operations and clustering at the output end of the NeRF. These improvements are applicable to all NeRF-based models.