Can you generate a 3D model from a single photo ?
I've been working on a new algorithm that makes this possible,
harnessing the power of machine learning to automatically generate a compact 3D model from a single image.
I use a variety of classical and modern computer science algorithms to achieve this. The first step is to pre-process the image and detect basic signatures,
then use a machine learning trained classifier to select the best candidate geometry, then recombine the top matches using a scoring system.
This multi-stage process reduces the initial vast number of combinations so a model can be computed quickly.
Note, these 3D models are not NeRFs nor Gaussian Splats - Neural Radiance Fields are a great innovation, but they output a set of neural weights, not a standard 3D model format.
In contrast the 'Quato' algorithm takes a single photo as input, and produces as output a single .obj file containing textured quads,
which can be loaded and displayed in almost any 3D CAD program, such as Sketchup, Revit or Blender.
The models produced are very compact, when compared to meshes produced from photogrammetry or LiDAR scanning.
Results are promising, as you can see below, although Im currently limited by access to GPU compute time for model training.
The examples below show the source photo next to the computed 3D model, which you can navigate around if you have a modern browser with webgl.