I just saw this brilliant work by NVIDIA where the auto-decimate complex leafy tree models from >5bn triangles down to 20M using some computer vision black magic. Since photogrammetry outputs are notoriously crappy and too high poly count, I was wondering if a method like this might be usable as a tool to improve model outputs, but this is far outside of my skillset to I’m not equipped to evaluate the technical constraints of this method or whether or not it would be suitable for use with photogrammetry outputs.
2 minute papers video: NVIDIA’s New Technique: Beautiful Models For Less! 🌲 - YouTube
Project page: Appearance-Driven Automatic 3D Model Simplification | Research
Code on github: GitHub - NVlabs/nvdiffmodeling: Differentiable rasterization applied to 3D model simplification tasks