Virtualized Point Cloud Rendering.

Area of Science:

• Computer Graphics
• Geospatial Technology
• Remote Sensing

Background:

• LiDAR and other remote sensing technologies generate massive point clouds, often exceeding GPU memory capacity.
• Existing Level of Detail (LoD) techniques for managing large point clouds are computationally intensive and time-consuming to build.
• Processing and rendering extremely large point clouds presents significant challenges for storage and real-time performance.

Purpose of the Study:

• To propose a novel GPU-driven culling system for efficient processing and rendering of arbitrarily large point clouds.
• To develop a system that maintains a low memory footprint on both CPU and GPU.
• To enable real-time manipulation of massive point clouds on commodity hardware.

Main Methods:

• Points are organized into groups sorted using Hilbert encoding, avoiding issues found in Morton curves.
• A GPU-driven culling system determines visible points per frame, transferring data on demand.
• Hole filling is implemented to address gaps caused by data density and LoD techniques.

Main Results:

• The system successfully manipulated point clouds up to 18 billion points, achieving an average of 80 FPS.
• Real-time performance was maintained without perceptible loss in visual quality.
• The method demonstrated superior capability in handling significantly larger point clouds compared to state-of-the-art approaches.

Conclusions:

• The proposed GPU-driven culling system offers an efficient solution for real-time processing of massive point clouds.
• The Hilbert encoding approach and on-demand data transfer enable manipulation of huge datasets on low-memory hardware.
• The system provides a scalable and effective method for large-scale geospatial data visualization and analysis.