Pointcept/Pointcept

What it solves

Pointcept is a flexible codebase designed for point cloud perception research, providing a unified framework for learning 3D representations and performing perception tasks like semantic and instance segmentation.

How it works

It integrates a wide array of state-of-the-art 3D backbones, pre-training frameworks, and dataset handlers. The library supports various architectures including Point Transformer (V1, V2, V3), OA-CNNs, and Swin3D, and implements several self-supervised pre-training methods such as Utonia, Concerto, and Sonata. It provides standardized preprocessing pipelines for major 3D datasets like ScanNet, S3DIS, and SemanticKITTI.

Who it’s for

It is primarily intended for researchers and developers working on 3D computer vision, point cloud analysis, and embodied intelligence, allowing them to easily experiment with different backbones and pre-training strategies.

Highlights

• Extensive Model Zoo: Includes official implementations of Point Transformer V1-V3, OA-CNNs, and other high-performance backbones.
• Advanced Pre-training: Supports cutting-edge self-supervised learning frameworks like Utonia, Concerto, and Sonata.
• Broad Dataset Support: Built-in preprocessing for a vast range of indoor and outdoor datasets including ScanNet, S3DIS, Waymo, and nuScenes.
• Hardware Flexibility: Supports both NVIDIA GPUs (CUDA) and AMD GPUs (ROCm).
• Integration: Compatible with wandb for experiment tracking.