Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, Jonathan Li
A comprehensive survey of 3D neural field representation research over five years, from the advent of NeRF to Gaussian Splatting.
NeRF enabled novel view synthesis by learning implicit neural representations of 3D scenes, but suffered from slow training and rendering speeds and high memory usage. Gaussian Splatting improved this to real-time levels, yet a systematic comparison of the strengths, weaknesses, and applications of both approaches was lacking.
We collected NeRF and neural field papers published from 2020-2025, dividing them into pre- and post-Gaussian Splatting eras, and proposed an architecture- and application-based taxonomy. We explain the theoretical background of NeRF and its differentiable volume rendering training process, and compare key datasets and performance benchmarks.
We systematically organize the evolution of NeRF and Gaussian Splatting, analyzing the strengths and weaknesses of each method. We also present their application status in various fields such as robotics, urban mapping, autonomous driving, and VR/AR, providing a useful reference for future researchers.