Abstract:  Image-based view synthesis techniques are widely applied to both computer graphics and computer vision. One of the key issues is how to use the information from the input image to represent a 3D model or scene. Recently, with the proposal of neural radiance fields (NeRF), a large number of research works based on this representation have further enhanced and extended the method, and achieved the expected accuracy and efficiency. This type of research can be broadly classified into two categories by purposes: the analysis and improvement of NeRF itself, and the extensions based on the NeRF framework. Methods of the first category have analyzed the theoretical properties and shortcomings of the NeRF representation and proposed some strategies for performance improvement, including the synthesis accuracy, rendering efficiency, and model generalizability. The second type of works are based on the NeRF framework and have extended the algorithm to solve more complex problems, including view synthesis using unconstrained images, view synthesis with relighting, and view synthesis for dynamic scenes. After outlining the background of the proposal of NeRF, other related works based on it were discussed and analyzed in this paper according to the classification mentioned above. Finally, the challenges and prospects were presented concerning the development of NeRF-based approaches. 

Key words:  image-based rendering, view synthesis, neural radiance fields, neural rendering, deep learning