Please note: This seminar will take place in DC 2310.
Letao Chen, PhD student
Department of Computer Science, University of Southern California
How can we refine sparse signed distance samples while ensuring that they remain consistent with a geometrically realizable surface? In this talk, I will present a method for interpolating exact signed distance function (SDF) data that guarantees consistency both with the input samples and among newly generated values. Our approach translates the theoretical properties of exact SDFs into hard geometric constraints and uses an efficient greedy algorithm, accelerated by parallel GPU preprocessing, to construct consistent samples. I will also show the usefulness through the following three applications: upsampling SDF data without access to ground truth, reconstructing detailed surfaces from coarse samples using global information, and repairing pseudo-SDFs produced by operations such as constructive solid geometry (CSG) Booleans.