Square-LFRF: high-speed light field 3D source generation using anti-aliasing virtual views synthesis based on neural...

Area of Science:

• Computer Vision
• Computer Graphics
• Display Technology

Background:

• Integral imaging enables glasses-free 3D perception by generating parallax in two dimensions.
• Light field 3D sources, crucial for these displays, are typically merged from multiple parallax images.
• Generating these sources faces limitations in speed and quality due to the increasing number of viewpoints.

Purpose of the Study:

• To develop a high-speed, high-fidelity method for generating light field 3D sources.
• To enhance virtual view synthesis for improved 3D display quality.
• To overcome the limitations of existing methods in generating complex light field data.

Main Methods:

• Proposed Square-LFRF method utilizing neural radiance fields (NeRF) for virtual view synthesis.
• Employed square pyramid frustum casting and recalibrated sampling integrated position encoding.
• Implemented cubic projection for network size and training time reduction.
• Introduced a pixel culling method to optimize rendering by eliminating redundant rays.

Main Results:

• Square-LFRF achieved an 8% reduction in average relative error rates compared to Mip-NeRF on the Blender dataset.
• Network size was reduced to 72.3 MB and training time to 10 minutes.
• The pixel culling method accelerated light field 3D source generation by 70% compared to cutoff-NeRF.

Conclusions:

• Square-LFRF offers a significant advancement in generating high-quality, anti-aliasing virtual views for light field 3D sources.
• The method demonstrates substantial improvements in speed and efficiency for NeRF-based 3D source generation.
• Square-LFRF paves the way for more practical and efficient glasses-free 3D display technologies.