Neural Radiance Fields (NeRF) are emerging as a viable option for 3D model capture. NeRF allows the creation of 3D meshes using a combination of sparse 2D photos and deep learning. Time savings can be realized by creating 3D models using this technique, allowing faster training and simulation content development. However, no new technology is without challenges.

The initial hurdle involves selecting suitable NeRF processing software and configuring it. Our study identified three NeRF processing candidates: Luma.ai, Instant NeRF, and NeRF Studios. For our investigation, we opted for Instant NeRF due to its compatibility with local desktop processing and fewer issues with its viewer than NeRF Studios' web viewer.

The second challenge lies in creating a robust NeRF. Technical limitations exist, particularly for those with a background in photogrammetry. Image sets must exceed 50 but not exceed 200 images, as an excess causes stutter in NeRF viewers, while too few results in suboptimal NeRF outcomes. We opted for 4k photogrammetry image sets over video to prevent image blurring. A high shutter speed camera with a fixed aperture is ideal for video applications.

The third challenge is ensuring that images are captured with the needs of NeRF in mind. There are several factors that photographers need to be aware of when capturing their NeRF Subject that people coming from photogrammetry may need to be made aware of. These dependencies involve factors such as lighting conditions, orientation of the subject, orientation of the camera, etc.

This paper discusses the merits and drawbacks of NeRF versus Photogrammetry in the context of 3D scanning and capture. Technical challenges, such as those described in preceding paragraphs, and our approach to overcoming those challenges are discussed. The paper provides visual examples to compare NeRF and Photogrammetry Mesh Exports and explore the potential gains of using NeRF and Photogrammetry together.

Hunter Stinson

Affiliation

Refine Results

Using VR to Validate and Visualize MBSE-Designed Interfaces

More Than a Kid’s Toy: Using NeRF to Create 3D Models