How to turn a collection of 2D images into a 3D model? One process is called “multiview stereo reconstruction” which has been using famous device “the Stanford spherical gantry” to take photos of an object from points on a spherical shell surrounding to build 3D model.
However often even with such nice data, the multiview stereo problem is still ill-posed. mikolalysenkowanted to reconstruct volumetric graphics from 8-bit sprites so he built a much simpler editor called space carving. “The underlying idea behind space carving is to reformulate the multiview stereo matching problem as a fairly simple geometry problem”.
if we suppose that our object has some predetermined material properties (ie is Lambertian and non-transparent) and that all our lighting is fixed, then we can check consistency by just rendering our object from each view and comparing the images pixel-by-pixel.
It is not the most accurate multiview stereo algorithm, but it reformulates the multiview stereo matching problem simply by inputting colored pixels onto 6 2D grids.
To keep things simple, let us suppose that we view each sprite directly along the 6 cardinal directions (ie top, bottom, front, back, left, right) and that our cameras are all orthographic projections:
Putting this all together with the above sweep line method gives us a pretty simple way to edit sprites directly.
You can try the WebGL shape carving editor here.
mikolalysenko sent the 3D models off to Shapeways for 3D printing. The cool miniatures are about 5cm or 2in tall and now available at this Shapeways shop.
Posted in 3D Software