Less is More: Skim Transformer for Light Field Image Super-resolution
New 'Skim Transformer' architecture achieves state-of-the-art results by intelligently ignoring redundant image data.
Deep Dive
A research team led by Zeke Zexi Hu introduces SkimLFSR, a novel network for light field image super-resolution. Built on a 'Skim Transformer' architecture, it uses only 67% of the parameters of prior leading methods while achieving superior performance, beating the best existing method by 0.63 dB PSNR at 2x upscaling. It works by selectively processing subsets of sub-aperture images to avoid computational waste on redundant data.
Why It Matters
Enables higher-quality 3D/VR content creation and medical imaging with significantly less computational cost.