Learning Sheared EPI Structure for Light Field Reconstruction.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Light field reconstruction aims to synthesize new views from captured light fields.
• Existing methods often rely on depth information, which can be challenging to acquire accurately.
• Epipolar plane images (EPIs) offer a structured representation of light field data.

Purpose of the Study:

• To develop a learning-based light field reconstruction method that eliminates the need for explicit depth estimation.
• To leverage the structural properties of sheared EPIs for accurate view synthesis.
• To demonstrate the versatility and performance of the proposed approach across various datasets.

Main Methods:

• A convolutional neural network (CNN) is trained to evaluate sheared EPIs.
• The CNN learns to identify the correct shear value corresponding to the depth of image patches.
• Sheared EPIs are fused based on scores generated by the CNN, without requiring depth maps.

Main Results:

• The proposed method achieves high performance in light field reconstruction on synthetic, real-world, and microscopic datasets.
• The network successfully learns to infer depth cues implicitly from sheared EPIs.
• Demonstrated capability for novel view synthesis and potential for depth inference.

Conclusions:

• The presented learning-based approach offers an effective and depth-agnostic solution for light field reconstruction.
• The method's reliance on sheared EPI structure simplifies the reconstruction pipeline.
• This work opens avenues for improved light field rendering and depth estimation techniques.