One repo, three papers: tiny models for giant point clouds

What it does

This is the official PyTorch implementation for three related methods: Superpoint Transformer (ICCV'23), SuperCluster (3DV'24 Oral), and EZ-SP (ICRA'26). All three tackle semantic or panoptic segmentation of large 3D point clouds by first compressing millions of raw points into a hierarchy of semantically coherent superpoints. Instead of brute-forcing attention across every single coordinate, the models run lightweight transformers or graph clustering on these condensed clusters.

The interesting bit

SuperCluster reframes panoptic segmentation as a graph optimization problem: the network predicts node and edge parameters, and a fast parallelized solver turns those into object instances at inference time, dodging the usual matching headaches during training. EZ-SP later replaces the original CPU-bound partitioning with a learnable GPU graph clustering pipeline that embeds points so different semantic classes naturally drift apart before clustering.

Key highlights

• Absurdly small footprints: SPT and SuperCluster weigh in at 212k parameters—roughly 1/200th the size of PointNeXt—while EZ-SP stays lean at 392k.
• Single-GPU scalability: SuperCluster digests an 18-million-point, 7.8 km² tile in 10.1 seconds; SPT trains on S3DIS in three hours on one GPU.
• EZ-SP claims a 72× end-to-end speedup over PTv3 and a 5.3× speedup over the original SPT by moving superpoint partitioning to the GPU.
• The codebase spans three tasks—semantic segmentation, panoptic segmentation, and fast learnable partitioning—under one Hydra-and-Lightning configuration roof.
• The core GPU graph algorithms are also available separately in the torch-graph-components library.

Caveats

• The README is truncated mid-sentence in the EZ-SP update section, so the full details of the latest release are partially missing.
• The authors explicitly note that if you only need the fast graph connected components or partitioning algorithms, you should use the standalone torch-graph-components library instead of this full segmentation framework.
• Every benchmark number in the README is drawn from S3DIS, KITTI-360, DALES, or ScanNet; no other datasets are discussed.

Verdict

Computer vision researchers and robotics engineers working with large-scale LiDAR or indoor scans who need off-the-shelf semantic or panoptic segmentation should start here. If you are looking for a general-purpose point cloud utility or need support for arbitrary datasets without heavy reconfiguration, this is probably not your tool.