RealSLF and FlexiDim: Towards Practical Spectral Light Field Imaging

Spectral light field (SLF) imaging simultaneously captures angular and spectral information, providing richer scene representations than conventional RGB imaging and enabling a wide range of scientific and industrial applications. However, current SLF research faces two key challenges: the lack of real-world datasets with high spatial, angular, and spectral resolutions, and the absence of efficient high-dimensional learning frameworks suitable for resource-constrained environments. To address these limitations, we present RealSLF, a large-scale, real-world SLF dataset featuring a wide range of spectral bands, dense viewpoints, and a scene count nine times greater than existing datasets. In addition, we develop a geometric correction pipeline that reduces acquisition misalignment, enhancing cross-view consistency. On the algorithmic front, we propose FlexiDim, a flexible and resource-affordable learning framework tailored for high-resolution SLF reconstruction under practical computational constraints. We evaluate its effectiveness by integrating it with various state-of-the-art super-resolution modules. Finally, we build a prototype RGB light field system based on a microlens array, demonstrating the practical feasibility of capturing high-resolution SLF data without additional optical components, and offering a scalable solution for SLF imaging.