Neural Radiance Field Reconstruction in Complex Scenes Without Camera Parameter Dependency

Neural Radiance Field (NeRF), which has been introduced since 2020, enables realistic 3D reconstruction from calibrated images. NeRF inspired global research into Neural Volumetric Rendering (NVR). In this work, we will first explore NVR's foundations, its ties to image-based rendering and inverse graphics, and then evaluate three methods: NeRF, 3D Gaussian Splatting and Gaussian Shader. As a result, 3D Gaussian Splatting excels in rendering speed, while NeRF offers accessible 3D training. In addition, Gaussian Shader excels in handling reflective surfaces but struggles with complex scenes, where the other methods prove more robust.