by Alice Mehalek, Marcus Vidaurri, and Zhecheng Wang
UV mapping or UV parameterization is the process of mapping a 3D surface to a 2D plane. A UV map assigns every point on the surface to a point on the plane, so that a 2D texture can be applied to a 3D object. A simple example is the representation of the Earth as a 2D map. Because there is no perfect way to project the surface of a globe onto a plane, many different map projections have been made, each with different types and degrees of distortion. Some of these maps use “cuts” or discontinuities to fragment the Earth’s surface.
In computer graphics, the typical method of UV mapping is by representing a 3D object as a triangle mesh and explicitly assigning each vertex on the mesh to a point on the UV plane. This method is slow and must be repeated often, as the same texture map can’t be used by different meshes of the same object.
For any kind of parametrization or UV mapping, a good UV map must be injective and should be conformal (preserving angles) while having few cuts. Ideally it should also be isometric (preserve relative areas). In general, however, more cuts are needed to achieve less distortion.
Our research mentor, Marco Tarini, developed the method of Volume-encoded UV mapping. In this method, a surface is represented by parametric functions and each point on the surface is mapped to a UV position as a function of its 3D position. This is done by enclosing the surface or portion of the surface in a unit cube, or voxel, and assigning UV coordinates to each of the cube’s eight vertices. All other points on the surface can be mapped by trilinear interpolation. Volume-encoded parametrization has the advantage of only needing to store eight sets of UV coordinates per voxel, instead of unique locations of many mesh vertices, and can be applied to different types of surface representations, not just meshes.
We spent the first week of our research project learning about volume-encoded parametrization by exploring the 2D equivalent: mapping 2D curves to a one-dimensional line, u. Given a curve enclosed within a unit square, our task was to find the u-value for each corner of the square that optimized injectivity, conformality, and orthogonality. We did this using a Least Squares method to solve a linear system consisting of a set of constraints applied to points on the surface. All other points on the curve could then be mapped to u by bilinear interpolation.
In Week 2 of the project, we moved on to 3D and created UV maps for portions of a sphere, using a similar method to the 2D case. Some of the questions we wanted to answer were: For what types of surfaces is volume-encoded parametrization possible? At what level of shape complexity is it necessary to introduce cuts, or split up the surface into smaller voxels? In the 2D case, we found that injectivity could be preserved when mapping curves less than half a circle, but there were overlaps for curves greater than a semicircle.
One dimension up, when we go into the 3D space, we found that uniform sampling on the quarter sphere was challenging. Sampling uniformly in the 2D parametric space will result in a distorted sampling that becomes denser when getting closer to the north pole of the sphere. We tried three different approaches: rewriting the parametric equations, sampling in a unit-sphere, and sampling in the 3D target space and then projecting the sample points back to the surface. Unfortunately, all three methods only worked with a certain parametric sphere equation.
In conclusion, over the two weeks, we designed and tested a volume-encoded least-squares conformal mapping in both 2D and 3D. In the future, we plan to rewrite the code in C++ and run more experiments.