3D used to mean C++; it doesn’t anymore.

This talk is about how that ecosystem actually fits together, using one running example: a raw LiDAR scan of a street, turned into labelled objects (ground, buildings, vegetation, street furniture) using Python.

We’ll cover:

• The map of the ecosystem, and how to pick: Open3D, PyVista, trimesh, laspy, SciPy’s KDTree, PyTorch3D. What each is for, and which combinations are worth the trouble.

• The performance cliff. These libraries are thin bindings over C++ kernels, so one Python-level for loop over your points can cost you two orders of magnitude. Where the boundary sits, and how to stay on the right side of it.

• How to fit millions of points into finite RAM (chunked reads, and downsampling,etc)

You’ll leave with a mental model of the 3D Python stack, a working end-to-end pipeline to copy, and the specific gotchas that cost me weeks.