Definition: NeRF (Neural Radiance Fields) is a method that uses a series of photos to train a neural network that describes a scene as a three-dimensional light field. From it, you can calculate photorealistic views from angles that were never photographed ("novel view synthesis"). Ben Mildenhall and colleagues presented the method in 2020 at the ECCV conference.

English: Neural Radiance Fields (NeRF)
German: Neuronale Strahlungsfelder

How It Works

  1. Take photos: Dozens to hundreds of images of an object or room from different positions.
  2. Determine the camera positions: Software calculates for each photo where the camera was and where it pointed.
  3. Train the network: For each point in space and each viewing direction, the neural network learns how dense the scene is there and which color it shows.
  4. Render new views: For each pixel, rays are traced through the scene, and the values along the ray are summed to one color (volume rendering).

Because the color depends on the viewing direction, NeRF often shows reflective surfaces, gloss and fine structures more convincingly than classic photogrammetry. The original paper is freely available (arXiv 2003.08934).

NeRF and Gaussian splatting: In 2023, Kerbl et al. presented 3D Gaussian splatting. It does not describe the scene with a neural network, but with millions of small, colored ellipsoids, and it can be shown in real time. In practice, Gaussian splatting has replaced NeRF for many applications. For glTF, there is the ratified Khronos extension KHR_gaussian_splatting.

Limitations: NeRF does not give a classic 3D model made of triangles, but a light field. You cannot read dimensions, assemblies and materials directly. Rendering requires much computing power, and the scene must be static during the capture.

Importance in Mechanical Engineering

  • As-built documentation: You can capture production halls, legacy plants or construction sites photorealistically from photos or drone videos when no CAD data exists.
  • Remote acceptance and service: The customer and the service technician look at a captured plant from any angle, without being on site.
  • Addition to CAD: The real environment comes from the capture, the new machine as an exact CAD-based model, for example for layout planning or a digital twin.

Frequently Asked Questions (FAQ)

Does NeRF replace CAD data?

No. NeRF shows how something looks, not how it is designed. For configurators, exploded views or dimensions, the CAD-based model stays the basis.

Do I need special hardware for the capture?

For the capture, a camera or smartphone is sufficient. The training needs a powerful GPU or a cloud service.

What is the difference from LiDAR?

LiDAR measures distances directly and gives precise point clouds. NeRF derives the scene only from photos. It is visually more realistic, but less dimensionally accurate.