3D Gaussian splat trained from a DJI Manifold 3 Vertical 3D Smart Capture dataset captured by Michael Smithkey / SkyGator.
In this workflow, the Manifold 3 is used with the DJI Matrice 4E to automate a close-range vertical capture of a building facade, producing a very high-resolution dataset at roughly 1.5 mm/px GSD.
The goal of this test was to see how LichtFeld Studio handles an unusually close, high-GSD image set that was originally captured for standard photogrammetry / mesh reconstruction rather than specifically for 3D Gaussian Splatting.
The mesh result from Mike’s original video keeps stronger close-up detail on the flat building geometry, while the splat handles elements like glass, reflections, vegetation, and thin structures better. The surrounding context is more limited, since the dataset was focused mainly on the facade rather than capturing the environment from many viewpoints. Close-up facade details could likely be meaningfully improved with targeted densification on the building surface, and/or a higher splat count.
Mike’s original video shows the capture workflow and mesh result from the same dataset:
https://www.youtube.com/watch?v=mEqzlpuwGzI
Capture/data: Michael Smithkey / SkyGator
https://skygator.com/
Platform: DJI Matrice 4E + DJI Manifold 3
Processing: RealityScan + LichtFeld Studio
Alos Engineering
https://alos.no/