This arXiv preprint describes a system that builds textured triangle meshes from several RGB-D sensors in real time, with the whole pipeline running on the GPU.
Pipeline: TSDF fusion, Marching Cubes, noise removal, texture mapping, inpainting and seam correction all stay on the GPU, with no intermediate mesh or image data returned to the host.
Seams: Each face is textured from a single camera, so the authors estimate a per-camera low-frequency gain field and bake it into the atlas to handle exposure seams.
Results: On the CMU Panoptic dataset, the method is faster than the baselines at every tested voxel resolution and sensor count, with peak speedups above an order of magnitude over CUDA nvblox and CPU Open3D. Re-rendering quality approaches an offline textured-meshing baseline.
Status: Single author, accepted to WACV 2027, 20 pages including appendix.
Sign in to comment.
TL;DR of the paper, generated automatically.
This arXiv preprint describes a system that builds textured triangle meshes from several RGB-D sensors in real time, with the whole pipeline running on the GPU.