What it does

A complete, step-by-step pipeline for Semantic3D semantic-8 segmentation: load raw point clouds, downsample with Open3D’s voxel grid, train a PointNet++ model in TensorFlow, predict on sparse samples, then interpolate labels back to the full dense cloud via k-NN search. Think of it as a cookbook showing how Open3D’s C++-accelerated utilities can slot into a deep learning workflow without writing your own I/O or geometry code.

The interesting bit

The value isn’t the model—it’s the plumbing. Open3D handles the unglamorous but critical stages: efficient PCD I/O, voxel downsampling, and fast radius-nearest-neighbor interpolation to upsample sparse predictions. The README even includes a renamer.py for benchmark submission, which is either thoughtful or telling about how much of this work is bureaucratic formatting.

Key highlights

• Fork chain: original PointNet2 → mathieuorhan’s semantic fork → this Open3D integration layer
• Requires compiling custom TensorFlow ops (CUDA + CMake 3.8+) for grouping, interpolation, and sampling
• Full pipeline: raw TXT → PCD conversion → voxel downsampling → training → sparse prediction → k-NN interpolation → dense labels
• Published benchmark entry on semantic3d.net (method: PointNet2_Demo)
• Repository is archived; active development moved to isl-org/Open3D-ML

Caveats

• The “Moved to Open3D-ML” banner is the only content above the fold—this is explicitly a legacy reference
• Custom TF op compilation is a friction point; the README warns about CUDA/CMake version requirements but offers no troubleshooting guidance
• No performance numbers, training times, or hardware requirements listed

Verdict

Worth a quick read if you’re maintaining an old Open3D-based point cloud pipeline or want to see how geometry libraries integrate with TF. Everyone else should head straight to Open3D-ML, which presumably absorbed whatever lessons this demo taught.