Spatial-Angular Attention Network for Light Field Reconstruction.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional light field reconstruction methods require deep networks to capture view correspondences.
• Existing approaches struggle with non-local dependencies in high-dimensional light field data.

Purpose of the Study:

• To propose a novel spatial-angular attention network for end-to-end light field reconstruction.
• To enhance the reconstruction of high angular resolution light fields from sparse inputs.

Main Methods:

• Introduced a spatial-angular attention module to capture non-local correspondences across angular and spatial dimensions.
• Developed a multi-scale reconstruction structure for efficient non-local attention processing.
• Utilized a non-local attention mechanism adapted for high-dimensional light field data.

Main Results:

• The proposed network effectively perceives non-local correspondences in light fields.
• Achieved superior performance in reconstructing sparsely-sampled light fields.
• Demonstrated robustness in handling Non-Lambertian effects during reconstruction.

Conclusions:

• The spatial-angular attention network offers a significant advancement in light field reconstruction.
• The method provides an efficient and effective solution for high angular resolution light field generation.
• This approach addresses key limitations of current learning-based light field reconstruction techniques.