How to run Superpoint Transformer on my own 3D point cloud?

Follow the tutorial slides, notebook, and video provided; it requires preprocessing your data into the hierarchical superpoint structure, tuning partition parameters, and loading a pretrained checkpoint, which can be complex for unfamiliar datasets.

Superpoint Transformer vs PointNet for semantic segmentation?

SPT is more efficient and scalable for large scenes, with far fewer parameters and faster training, but PointNet might be simpler for small-scale tasks or beginners due to its straightforward architecture and wider community support.

What GPU memory do I need to train SuperCluster?

Default configs are for 32G GPUs, but 11G GPUs can be used with adjusted parameters like xy_tiling or voxel size, though this may increase training time and affect performance, as detailed in the troubleshooting section.

How to fix CUDA out-of-memory errors in this project?

Adjust datamodule parameters such as xy_tiling, max_num_nodes, or use gradient accumulation; the README includes a detailed table of memory-affecting settings and configs for 11G GPUs to mitigate issues.

Can I use EZ-SP without TorchSparse?

TorchSparse is optional but enables sparse convolutions for EZ-SP; without it, you might miss performance optimizations, though basic functionality may still work with adjusted models, as noted in the installation script.

What's the difference between SPT, SuperCluster, and EZ-SP?

SPT is for semantic segmentation using superpoint transformers, SuperCluster extends it for panoptic segmentation via graph clustering, and EZ-SP is a faster variant with learnable, GPU-accelerated partitioning, but all share core superpoint-based architectures.

How often does this project update with breaking changes?

Updates like the EZ-SP release in November 2025 introduced non-backward compatible changes, requiring environment reinstallation; the CHANGELOG documents these, suggesting a research-focused pace that may disrupt long-term projects.

GitHub repository — 3D Segmentation with Superpoints

What is GitHub repository?

Superpoint Transformer is an open-source PyTorch implementation of research models for 3D point cloud segmentation. It introduces the Superpoint Transformer for efficient semantic segmentation and SuperCluster for scalable panoptic segmentation, both based on hierarchical superpoint graphs. The framework solves the problem of processing large-scale 3D scenes with limited computational resources by using lightweight transformer architectures and graph clustering formulations.

Target Audience

Researchers and developers working on 3D computer vision, particularly those focused on point cloud segmentation for applications like autonomous driving, robotics, and geospatial analysis. It's suited for those needing efficient models for large-scale 3D data.

Value Proposition

Developers choose Superpoint Transformer for its exceptional efficiency—models train in hours on a single GPU with very few parameters—while achieving competitive accuracy on standard benchmarks. Its unique superpoint-based approach and scalable graph clustering formulation offer a principled alternative to dense 3D convolutions.

Official PyTorch implementation of Superpoint Transformer [ICCV'23], SuperCluster [3DV'24 Oral], and EZ-SP [ICRA'26]

Use Cases

Best For

Semantic segmentation of large indoor scenes (e.g., S3DIS, ScanNet)
Panoptic segmentation of outdoor LiDAR point clouds (e.g., KITTI-360, DALES)
Research on efficient 3D deep learning architectures
Applications requiring real-time or near-real-time 3D scene understanding
Projects with limited GPU memory but large 3D datasets
Developing graph-based methods for 3D point cloud analysis

Not Ideal For

Production teams needing stable, versioned APIs without breaking changes
Developers without Linux systems or high-memory GPUs (requires 64G RAM and NVIDIA GPUs)
Applications demanding plug-and-play deployment with minimal configuration
Projects where real-time inference on edge devices is critical, despite efficiency gains

Pros & Cons

Pros

Extreme Parameter Efficiency

Models have as few as 212k parameters, making them 200x smaller than PointNeXt and 40x smaller than Stratified Transformer, as highlighted in the README benchmarks.

Rapid Training and Inference

SPT trains on S3DIS in 3 hours on one GPU, with preprocessing 7x faster than Superpoint Graph, and SuperCluster processes 18M points in 10.1s, enabling large-scale processing.

Scalable Panoptic Segmentation

SuperCluster formulates panoptic segmentation as superpoint graph clustering, allowing it to handle massive scenes on a single GPU with fewer than 1M parameters, per the paper results.

Strong Benchmark Performance

Achieves state-of-the-art mIoU and PQ scores on datasets like S3DIS, KITTI-360, and DALES, with results backed by Papers with Code badges in the README.

Interactive Visualization Tools

Includes tools for creating shareable HTML visualizations of 3D segmentation results, enhancing interpretability and collaboration, as shown in the notebooks and media.

Cons

Complex Installation and Setup

Requires specific hardware (64G RAM, Linux OS), and the install.sh script has optional dependencies like TorchSparse, making it less accessible for casual users.

Backward Compatibility Breaks

The README explicitly notes non-backward compatible changes in updates, such as the EZ-SP release, forcing users to reinstall environments and reprocess datasets.

Steep Learning Curve for Customization

Parameterizing superpoint partitions and graph clustering requires deep understanding, as admitted in the tutorials, which may deter non-experts from adapting it to new data.

Limited Out-of-the-Box Support

Focuses on research benchmarks; using custom data involves significant preprocessing and hyperparameter tuning, with less documentation for production pipelines compared to broader frameworks.

Frequently Asked Questions

What is GitHub repository?

Target Audience

Value Proposition

Use Cases

Best For

Semantic segmentation of large indoor scenes (e.g., S3DIS, ScanNet)
Panoptic segmentation of outdoor LiDAR point clouds (e.g., KITTI-360, DALES)
Research on efficient 3D deep learning architectures
Applications requiring real-time or near-real-time 3D scene understanding
Projects with limited GPU memory but large 3D datasets
Developing graph-based methods for 3D point cloud analysis

Not Ideal For

Production teams needing stable, versioned APIs without breaking changes
Developers without Linux systems or high-memory GPUs (requires 64G RAM and NVIDIA GPUs)
Applications demanding plug-and-play deployment with minimal configuration
Projects where real-time inference on edge devices is critical, despite efficiency gains

Pros & Cons

Pros

Extreme Parameter Efficiency

Models have as few as 212k parameters, making them 200x smaller than PointNeXt and 40x smaller than Stratified Transformer, as highlighted in the README benchmarks.

Rapid Training and Inference

SPT trains on S3DIS in 3 hours on one GPU, with preprocessing 7x faster than Superpoint Graph, and SuperCluster processes 18M points in 10.1s, enabling large-scale processing.

Scalable Panoptic Segmentation

SuperCluster formulates panoptic segmentation as superpoint graph clustering, allowing it to handle massive scenes on a single GPU with fewer than 1M parameters, per the paper results.

Strong Benchmark Performance

Achieves state-of-the-art mIoU and PQ scores on datasets like S3DIS, KITTI-360, and DALES, with results backed by Papers with Code badges in the README.

Interactive Visualization Tools

Includes tools for creating shareable HTML visualizations of 3D segmentation results, enhancing interpretability and collaboration, as shown in the notebooks and media.

Cons

Complex Installation and Setup

Requires specific hardware (64G RAM, Linux OS), and the install.sh script has optional dependencies like TorchSparse, making it less accessible for casual users.

Backward Compatibility Breaks

The README explicitly notes non-backward compatible changes in updates, such as the EZ-SP release, forcing users to reinstall environments and reprocess datasets.

Steep Learning Curve for Customization

Parameterizing superpoint partitions and graph clustering requires deep understanding, as admitted in the tutorials, which may deter non-experts from adapting it to new data.

Limited Out-of-the-Box Support

Focuses on research benchmarks; using custom data involves significant preprocessing and hyperparameter tuning, with less documentation for production pipelines compared to broader frameworks.

Frequently Asked Questions

GitHub repository

What is GitHub repository?

Overview

Use Cases

Best For

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions

Related Projects

Found a gem we're missing?

GitHub repository

What is GitHub repository?

Overview

Use Cases

Best For

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions

Related Projects

Found a gem we're missing?