Pointcept/Pointcept

Pointcept: Perceive the world with sparse points, a codebase for point cloud perception research. Latest works: Utonia (ICML'26), Concerto (NeurIPS'25), Sonata (CVPR'25 Highlight), PTv3 (CVPR'24 Oral)

What it solves

Pointcept is a research codebase designed to simplify and accelerate the development of 3D point cloud perception models. It provides a unified, flexible framework that integrates state-of-the-art architectures, pre-training techniques, and datasets, allowing researchers to easily experiment with and benchmark new methods for tasks like semantic and instance segmentation.

How it works

The framework acts as a central hub for various 3D deep learning backbones and pre-training strategies. It includes standardized data preprocessing pipelines for major 3D datasets (such as ScanNet, S3DIS, and nuScenes), enabling users to swap between different models and training configurations with minimal friction. It supports modern training workflows, including integration with tools like Weights & Biases for experiment tracking.

Who it’s for

This project is intended for researchers and developers working in 3D computer vision and point cloud analysis. It is best suited for those who need a reliable, modular environment to implement, train, and evaluate cutting-edge 3D perception algorithms.

Highlights

  • Includes official implementations for high-impact research papers like Utonia, Concerto, Sonata, and Point Transformer V3.
  • Supports a wide range of backbones, including various SparseUNet variants, Point Transformers, and Swin3D.
  • Provides comprehensive data preprocessing scripts for numerous indoor and outdoor 3D datasets.
  • Offers flexible installation options via Conda or pre-built Docker images for consistent development environments.
  • Facilitates large-scale representation learning with support for multi-dataset pre-training frameworks.

Related

  • Project
    Pointcept/Concerto

    A joint 2D-3D self-supervised pre-trained Point Transformer V3 model for extracting high-quality spatial representations from 3D point clouds.

  • Project
    torch-points3d/torch-points3d

    A deep learning framework for point cloud analysis that provides a high-level API and a library of state-of-the-art models for 3D classification, segmentation, and detection.

  • Project
    prs-eth/LitePT

    LitePT is a lightweight, high-performance 3D point cloud architecture that reduces parameter count and memory footprint while maintaining state-of-the-art accuracy for 3D vision tasks.

  • Project
    guochengqian/openpoints

    A library for benchmarking and reproducing point-based methods for point cloud understanding, providing a standardized engine for models like PointNet and PointNeXt.

  • Project
    NVIDIA-AI-IOT/Lidar_AI_Solution

    A collection of highly optimized CUDA and TensorRT implementations for 3D Lidar perception, speeding up sparse convolutions and sensor fusion for self-driving systems.