Neural Subspaces for Light Fields.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Neural light field representation offers compression but lacks optimization for composite structures.
• A unified network design hinders rapid transmission and decoding of light field data.

Purpose of the Study:

• To develop a novel framework for compact light field representation using neural subspaces.
• To improve compression efficiency and reconstruction quality for light fields.

Main Methods:

• Utilizing multiple small neural networks, each learning a neural subspace for specific light field segments.
• Implementing an adaptive weight sharing strategy among networks for parameter efficiency.
• Employing a soft-classification technique to enhance color prediction accuracy.

Main Results:

• The proposed method achieves superior light field reconstruction compared to existing approaches.
• Demonstrated effectiveness on diverse light field scenes, including those with irregular viewpoints and dynamic content.
• Achieved improved parameter efficiency through adaptive weight sharing.

Conclusions:

• Neural subspaces offer a more optimized approach to light field compression than unified neural networks.
• The framework supports efficient encoding and decoding for complex and dynamic light field data.
• The method shows promise for real-world applications requiring high-fidelity light field representation.