Reproduce “A Unified Learning Based Framework for Light Field Reconstruction from Coded Projections” on OpenResearch

About this paper

Anil Kumar Vadathya, Sharath Girish, Kaushik Mitra

Researchers from Rice University, University of Maryland, and IIT Madras developed a unified deep learning framework for light field reconstruction from various coded projections, including a novel focus-defocus scheme requiring no specialized hardware. This approach achieves superior or competitive performance compared to state-of-the-art methods, demonstrating high-quality light field capture from conventional cameras.

Key insights

• Explicitly estimating a full disparity field, rather than a single depth map, from coded projections significantly improves the accuracy and interpretability of light field reconstruction.
• The use of multiscale context aggregation and dilated convolutions in DisparityNet is essential for resolving depth ambiguities and improving disparity map quality in challenging regions.
• High-quality light field capture can be achieved from a simple focus-defocus image pair using a conventional camera, democratizing access to LF imaging without specialized hardware.

Summary by alphaXiv.