LRFNet: Learning Light Field Reconstruction via a Large Receptive Field Network

Densely sampled light fields are powerful tools for applications such as post-capture refocusing and virtual reality, but acquiring such data remains costly and technically demanding. While existing reconstruction methods have shown promise, they often succeed only in small-baseline settings and struggle with larger disparities or real-time efficiency. Depth-based approaches are prone to artifacts due to imperfect depth estimates, while non-depth-based methods lack geometric accuracy, fail in occluded or textureless regions, and are typically computationally intensive. In this work, we provide a more effective disentanglement of spatial, angular, and epipolar representations for light field reconstruction. Through dedicated feature extractors and a residual-in-residual architecture enhanced with channel attention, our framework efficiently captures subpixel details and long-range dependencies while adaptively emphasizing the most informative cues. Rigorous ablation studies further highlight the critical role of epipolar feature interactions—an aspect previously overlooked in the literature. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach consistently surpasses state-of-the-art methods across small- and large-baseline scenarios, delivering higher reconstruction quality while maintaining competitive efficiency.