This post compares Gaussian splats to photogrammetry meshes and point clouds for complex geometries like industrial scenes with piping, hardware, and equipment. While photogrammetry works well for large, broad surfaces like rooms or buildings, it struggles with complex geometries where Gaussian splats excel.
To demonstrate this, a local park playground provided a test location with convoluted spaces and railings similar to industrial complexity. Using PIX4D, which tracks scan coverage, the video shows the scan path through the playground. The process captured 691 images, illustrating how many photos are needed for decent Gaussian splats.
The PIX4D web app can display different scan representations. Point clouds, shown in the video, require loading massive numbers of points yet still produce a grainy effect. For 3D scenes requiring realism, point clouds aren’t well suited.
Photogrammetry produces textured 3D meshes, but for reasonably complex scenes, they have too many artifacts, what could be called the melted cheese look. Unless you’ve got large smooth surfaces, photogrammetry isn’t the answer.
Gaussian splats, with enough samples and appropriate processing, achieve high levels of realism. The video shows this in the guardrails and smooth surfaces. For scenes with fine grain detail complexity, Gaussian splats provide much better realism than photogrammetry meshes or point clouds, with better performance.
Gaussian splats offer additional advantages. You can scan objects individually or extract portions from a scene in 3D. This enables programmatic manipulation like animating stepping stones in a game through movement, scale, or colorization. You can also use realistic scenes as backdrops and add other forms, mixing splats or standard meshes for game or simulation generation.
What are Gaussian splats? They’re basically ellipsoids (3D ellipses) with combinations of color, transparency, and shape. The training algorithm takes images as goals and identifies the optimal set of ellipsoids with different colors and transparency that would recreate each image from particular viewpoints. It’s sort of like 3D painting with ellipsoids.
This isn’t a press-button solution. You’ll need to edit out portions with insufficient photo coverage and re-scan if necessary. Don’t expect a low-cost workflow. To achieve higher realism, you need pedantic scanning with good coverage from all angles, consistent lighting, and careful post-processing. Missing camera angles will produce poor results, so complete coverage is essential for quality output.