Photogrammetry vs NeRF: Key Differences in 3D Reconstruction Explained

Added:

3D Gaussian Splatting: Moving Beyond the NeRF vs. Photogrammetry Dichotomy

While debates in 3D reconstruction often frame the future as a choice between traditional photogrammetry (explicit triangle meshes) and NeRFs (implicit neural volumetric representations), the rapid emergence of 3D Gaussian Splatting (3DGS) presents a disruptive third paradigm. Critics of both NeRF and photogrammetry argue that neither perfectly balances rendering speed with visual quality. NeRFs suffer from slow training and rendering times due to costly neural network queries, while photogrammetry struggles with specular reflections and complex, thin geometries. 3D Gaussian Splatting bypasses these limitations by using explicit, differentiable 3D Gaussians that can be rasterized in real-time. This technique offers the high-fidelity, view-dependent visual quality of NeRFs but achieves rendering speeds exceeding 100 FPS at high resolutions without requiring heavy neural network evaluation. Consequently, 3DGS challenges the assumption that neural fields are the only viable path forward for highly realistic, real-time volumetric XR environments.

Fundamental concepts of computer vision, specifically how 2D images are projected from 3D space (camera intrinsics and extrinsics).
The basics of traditional 3D graphics representations, particularly polygon meshes, vertices, and texture mapping.
Core principles of Photogrammetry, such as Structure from Motion (SfM) and point cloud generation.
An introductory understanding of machine learning and neural networks, particularly how they function as continuous function approximators.
Advanced neural rendering techniques, such as 3D Gaussian Splatting, which offer faster real-time rendering alternatives.
Pipeline integration of NeRF and photogrammetry assets into game engines like Unreal Engine and Unity for XR (VR/AR) applications.
Methods for dynamic NeRFs (handling moving objects and temporal changes in a 3D scene rather than static environments).
Hybrid reconstruction workflows that combine the structural accuracy of photogrammetry with the view-dependent volumetric effects of NeRFs.
410 views5likes4:56@allthingsxrpodcast2299Original Release: 2023-04-08

NeRF (Neural Radiance Fields) is a volumetric representation that captures radiance fields including background, perspective, reflectance, and specularity without requiring manual processing, while photogrammetry creates mesh-based representations using triangles with textures mapped onto them; NeRF focuses on making scenes look correct to the viewer without caring about geometric accuracy, whereas photogrammetry attempts to place objects in their correct 3D positions, making NeRF better for backgrounds and photogrammetry better for interactive objects in XR applications.