LiDAR segmentation by pretending it's a camera

What it does

LiDAR-Bonnetal trains semantic segmentation models for autonomous-driving LiDAR scans. Instead of grappling with raw 3D points, it projects them into 2D range images—distance and intensity as pixels—then runs familiar CNN backbones (SqueezeSeg, DarkNet21/53) over that grid. Labels predicted on the 2D map get mapped back to the original point cloud. The repo contains the training pipeline and a zoo of pre-trained weights for the SemanticKITTI benchmark.

The interesting bit

The trick is the round-trip: spherical projection makes the problem look like image segmentation, which is well-solved and fast on GPUs. A k-NN post-processing step can clean up discretization artifacts where the re-projection muddles fine structure. The authors published this as RangeNet++ at IROS 2019.

Key highlights

• Pre-trained models for eight architecture variants, with and without CRF or k-NN post-processing
• Training code lives in /train; MIT-licensed framework
• Predictions provided for SemanticKITTI train, val, and test splits
• Companion dataset paper (SemanticKITTI, ICCV 2019) is the standard benchmark in the field

Caveats

• Repository is archived and unmaintained; no issues or PRs accepted
• Deployment pipeline was promised “soon” in the README and apparently never open-sourced
• Pre-trained model weights carry the copyright of their training dataset, so check license terms before commercial use

Verdict

Worth studying if you need a fast, proven baseline for LiDAR segmentation or want to understand range-image approaches. Skip it if you need active maintenance, a modern PyTorch rewrite, or a ready-to-run inference pipeline.