Moving least-squares reconstruction of large models with GPUs

Area of Science:

• Computer Vision
• Geometric Modeling
• High-Performance Computing

Background:

• Laser scanning generates massive point cloud datasets, necessitating efficient surface reconstruction methods.
• Existing out-of-core techniques often require substantial computational resources.
• Moving Least-Squares (MLS) is a powerful surface reconstruction technique but is computationally intensive.

Purpose of the Study:

• To present a novel GPU-accelerated, out-of-core implementation of the Moving Least-Squares (MLS) surface reconstruction technique.
• To enable efficient processing of extremely large point clouds from laser range scanning.
• To demonstrate the scalability and performance of the proposed method on multi-GPU systems.

Main Methods:

• Developed a GPU-accelerated, out-of-core MLS surface reconstruction algorithm.
• Implemented a spatial subdivision approach using cubic regions for independent processing.
• Utilized multi-GPU parallelization for enhanced computational efficiency.

Main Results:

• The GPU-accelerated implementation achieves an order-of-magnitude speedup compared to CPU-based methods on a single GPU.
• The out-of-core nature allows processing of billion-point datasets on a standard desktop PC.
• The method demonstrates excellent scalability across multiple GPUs (up to 8 GPUs).

Conclusions:

• The proposed GPU-accelerated MLS surface reconstruction is highly effective for large-scale laser scanned data.
• This approach significantly reduces processing time and computational requirements.
• It enables efficient 3D surface reconstruction from massive point clouds on accessible hardware.