NKSR

Why we included this project

NKSR adapts Neural Kernel Fields so they work at scene scale rather than only on single objects. Compact kernel functions feed memory-efficient sparse linear solvers, letting the method rebuild a clean, watertight mesh from millions of points in seconds and process very large scenes out-of-core. The gradient fitting solve tolerates noisy input, which matters when scans come from LiDAR, photogrammetry, or a phone rather than clean synthetic data. Teams doing digital twinning or SLAM post-processing who need to turn raw captured geometry into usable surface meshes will find this a solid, well-documented baseline to weigh against classic Poisson reconstruction.