prs-eth/LitePT

What it solves

LitePT addresses the high computational cost and memory requirements of 3D point cloud architectures, specifically targeting the inefficiencies of Point Transformer V3 (PTv3). It provides a high-performance alternative that is significantly lighter and faster without sacrificing accuracy.

How it works

LitePT is a lightweight 3D point cloud architecture available in multiple scales (Small, Base, Large). It utilizes a specialized component called PointROPE, which is provided in both a CUDA-optimized version for maximum performance and a pure PyTorch implementation for broader compatibility. The architecture is designed to be more efficient in terms of parameters and memory footprint while maintaining or exceeding the performance of state-of-the-art backbones.

Who it’s for

This project is for researchers and developers working with 3D computer vision, specifically those focusing on semantic segmentation, instance segmentation, and object detection within 3D point clouds.

Highlights

• High Efficiency: The LitePT-S variant has 3.6x fewer parameters, 2x faster runtime, and 2x lower memory usage compared to PTv3.
• Scalable Performance: Even the large variant (LitePT-L) remains more memory-efficient than PTv3 despite having more parameters.
• Versatile Applications: Supports semantic segmentation, instance segmentation, and object detection across benchmarks like NuScenes, Waymo, ScanNet, and Structured3D.
• Flexible Integration: Provides a standalone implementation and a demo script for easy integration into external projects.