Methods and strategies for improving the novel view synthesis quality of neural radiation field

In recent years, researchers have increasingly focused on the NeRF[1] in this regard. NeRF provides an accurate and simple method to represent 3D scenes useing an implicit function based on MLPand, and has achieved satisfactory rendering quality in 3D reconstruction tasks. Current efforts aim to extend the original NeRF to different situations, such as scene synthesis[2, 3], dynamic scenes[4, 5], large scene reconstruction[6, 7] or rapid convergence[8, 9], among others. Since NeRF was published in 2020, the NeRF paper has been cited more than thousands of times in the past three years. In addition, researchers have made numerous improvements to the NeRF technology. Some work have focused on optimizing the rendering speed of NeRF[10, 11], while others have explored different application scenarios[12, 13]. Futhermore, there have been efforts to extended NeRF for scene inpainting[14, 15], texture synthesis[16], handing complex scenes[17], and addressing more challenging problems.