Arbitrary-Scale Light Field Super-Resolution via Spatio-Angular Consistent Neural Epipolar Geometry Learning

Abstract Light field imaging, a promising technology in computer vision, captures both spatial and angular information of light rays. However, the inherent trade-off between spatial and angular resolution limits its applications. This paper introduces a novel spatio-angular consistent local neural epipolar geometry learning framework for arbitrary-scale light field super-resolution. By modelling light field epipolar plane images (EPIs) as continuous local implicit geometry functions, our approach enables simultaneous super-resolution across both angular and spatial domains. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of our model, achieving high geometric structure fidelity under arbitrary upscaling factors.