L3Fnet

L3Fnet restores low-light light field (LF) images using a deep neural network that enhances visual quality while preserving epipolar geometry for refocusing and depth estimation.

Key Features:

• Two-Stage Architecture: Stage-I encodes geometric information across all LF views and Stage-II reconstructs individual LF views using the encoded geometry.
• Epipolar Geometry Preservation: Maintains epipolar geometry across LF views to ensure coherent multi-view reconstruction and consistent depth cues.
• Adaptability to Low-Light Conditions: A pre-processing block adjusts the network response based on scene light level to handle lighting from moderately dim to near-zero lux.
• LF Dataset: Training data include four LF captures per scene (one optimal exposure and three progressively lower light levels) plus an L3F-wild subset of night-time captures with near-zero lux and no ground truth.
• Single-Frame Enhancement: Converts single-frame images into a pseudo-LF format to enable LF-based restoration for non-light-field inputs.

Scientific Applications:

• Computational Photography: Improves LF-based refocusing and denoising under low-light scenarios.
• Computer Vision: Provides enhanced multi-view inputs that preserve geometric consistency for downstream vision tasks.
• Depth Estimation and Refocusing: Preserves and enhances depth cues across views to support precise depth estimation and post-capture refocusing.
• Low-Light and Night-time Imaging Research: Enables analysis of LF data captured at near-zero lux for research without available ground truth.

Methodology:

Training uses a diverse LF dataset captured at multiple exposure levels; a pre-processing block adjusts for scene light level, Stage-I encodes geometric cues across views, and Stage-II reconstructs enhanced LF views.