Efficiency-Accuracy Trade-Off in Light Field Estimation with Cost Volume Construction and Aggregation.

Area of Science:

• Computer Vision
• Photogrammetry

Background:

• Light field images offer rich spatial and angular data for depth estimation, vital for environmental perception.
• High computational costs and memory demands often limit the practical application of light field depth estimation.
• Pruning light field data improves efficiency but reduces accuracy, particularly in low-texture or occluded areas.

Purpose of the Study:

• To develop a lightweight disparity estimation model for light field images.
• To enhance depth estimation accuracy, especially in textureless and occluded regions.
• To balance computational efficiency with high accuracy in depth estimation.

Main Methods:

• Combined absolute difference and correlation cost matching to build robust cost volumes.
• Developed a multi-scale disparity cost fusion architecture using 3D convolutions and a UNet-like structure.
• Integrated information across multiple depth scales for efficient cost volume fusion and completion.

Main Results:

• Achieved computational efficiency comparable to state-of-the-art efficient methods.
• Doubled the accuracy compared to existing efficient light field depth estimation techniques.
• Reached accuracy levels similar to the highest-accuracy methods with significantly improved computational performance.

Conclusions:

• The proposed lightweight model effectively enhances depth estimation accuracy in challenging regions.
• The method provides a superior balance between speed and accuracy for light field depth estimation.
• This approach offers a practical solution for real-time environmental perception using light fields.