Revisiting Light Field Rendering With Deep Anti-Aliasing Neural Network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Light field (LF) reconstruction faces challenges with large disparity and non-Lambertian effects.
• Existing methods often tackle these issues separately, lacking a unified approach.
• Aliasing is identified as the core problem in both large disparity and non-Lambertian rendering.

Purpose of the Study:

• To develop a unified deep learning framework for light field reconstruction.
• To address both large disparity and non-Lambertian challenges simultaneously.
• To improve light field view interpolation and extrapolation.

Main Methods:

• An anti-aliasing reconstruction framework is proposed, operating in the image domain.
• This framework is integrated into a deep neural network with an end-to-end trainable architecture.
• The network is trained using a diverse dataset including regular and unstructured light fields.

Main Results:

• The proposed deep learning pipeline demonstrates significant superiority over state-of-the-art methods.
• It effectively solves both large disparity and non-Lambertian challenges in LF reconstruction.
• The method also shows benefits for light field view extrapolation.

Conclusions:

• The developed deep learning approach offers a unified solution for complex light field reconstruction problems.
• Image-domain anti-aliasing within a neural network is a viable alternative to Fourier-domain methods.
• The framework enhances both interpolation and extrapolation capabilities for light fields.