Novel View Synthesis without Camera Parameters: A Survey

Recent advancements in the field of novel view synthesis have been extensively driven by the emergence of neural radiance fields (NeRFs) and 3D Gaussian splatting (3DGS). NeRF leverages a simple multilayer perceptron (MLP) and ray casting to create an implicit representation of a scene. Following NeRF, 3D Gaussian splatting was introduced, which utilizes point clouds and 3D Gaussians to create an explicit representation while also incorporating techniques that enable faster rendering capabilities. Subsequent developments have continued to enhance both the quality and rendering speed. However, the performance of methods based on NeRF and 3DGS is heavily influenced by the initial camera parameters. Camera intrinsics, which include the focal length, and camera extrinsics, which describe the camera’s pose, can either be precisely measured beforehand or estimated using structure from motion techniques such as COLMAP. Despite these methods, inaccuracies due to various factors can degrade the performance. To address these challenges and improve both the rendering speed and practicality, camera-parameter-free approaches are being actively researched. This survey focuses on such research within the domains of NeRF and 3DGS. Through this survey, we aim to introduce camera-parameter-free methods, thus providing motivation for further research in these areas.