Gaussian Splats with the DLP

At the 3D Digitization Lab at VT (official abbreviation DLP), we are sometimes limited in types of objects we can accurately digitize. Photogrammetry struggles to accurately capture shiny, translucent, grainy, and extremely dark subjects. However, Gaussian Splat technology has shown so far to be a potential fix for these struggles. I am currently pioneering a potential Gaussian Splat workflow that will work for the lab. This is an example of my research so far.

Specimen G227 is covered in semi-translucent cube mineral formations, and also has a grain-like texture over the entire surface. When our photogrammetry workflow was applied, Agisoft was confused by the semi-translucent cubes and bumpy grain, and in turn added texture to the smooth cubes and too much grain to the rest of the surface.

To start a Gaussian Splat workflow, I exported camera angles and images out of Agisoft and imported those into PostShot, where the trainer started making a splat of G227. After training, I cleaned up outlying splats. Comparing the photogrammetry workflow to the splat workflow, using Gaussian Splat technology allowed for a more accurate digitization of the specimen.

The first image is the Gaussian Splat, and the second is the model made with Photogrammetry.
_______________________________________________________

Made using Agisoft Metashape and Postshot