BigEPIT: Scaling EPIT for Light Field Image Super-Resolution

Abstract

Existing methods have been developed for light field (LF) image Super-Resolution (SR) and achieved continuously improved performance while suffering a significant performance drop when handling scenes with large disparity variations. EPIT [1] was proposed to mitigate the disparity issue through non-local spatial-angular correlation learning. However EPIT has limitations due to the limited scale of existing LF datasets and the presence of imbalanced LF disparity, especially the scarcity of large disparity. To address this issue, we present a series of strategies to scale EPIT, called BigEPIT, including compound model scaling, augmented data resampling, and a high-precision test scheme. Specifically, the compound scaling method simultaneously scales the depth and width of the model to better improve the model capability. The augmented resampling method employs varying sampling intervals during training data generation, rather than solely relying on the central region view. This approach mitigates issues related to disparity imbalance and overfitting. The patch-based test scheme is popular because of its small GPU memory footprint. The traditional zero padding method and window partition will destroy the LF disparity structure and degrade the performance. Moreover, we find a positive correlation between the performance and the patchsize. Therefore, we advocate a high-precision test scheme i.e., a full-size or larger patchsize without zero padding for testing wherever the GPU memory permits, to achieve superior results. Extensive experiments demonstrate the effectiveness of our proposed method, which ranked 1st place in the NTIRE 2024 Light Field Image Super-Resolution Challenge.