A Comprehensive Survey of Neural Radiance Fields for 3D Scene Reconstruction

Abstract

Three-dimensional (3D) scene reconstruction is a long-standing problem in
computer vision, with applications in augmented/virtual reality, autonomous robotics, and
computer graphics. Recently, Neural Radiance Fields (NeRFs) have emerged as a powerful
new paradigm for 3D representation and novel view synthesis. This literature review provides
an overview of both classical 3D reconstruction methods and modern neural rendering
approaches, with equal emphasis on each. We cover the foundations of multi-view 3D reconstruction
(structure-from-motion and multi-view stereo) alongside the formulation of NeRFs
and implicit volumetric representations. We then survey advances in neural rendering and
volumetric scene representations, including efficient NeRF variants (such as Instant Neural
Graphics Primitives and PlenOctrees) and techniques for real-time rendering. Integration
of NeRFs with SLAM and robotics is discussed, highlighting how neural representations
are being combined with simultaneous localization and mapping. Benchmark datasets and
evaluation metrics common to both domains are summarized. The review is organized in an
IEEE conference style, with sections on Introduction, Methodologies (classical and neural),
Discussion of current challenges, and Conclusion. A comprehensive reference list in IEEE
format is provided.