When Corruptions Disrupt Cross-View Geometry in Light Field Depth Estimation

Light field (LF) depth estimation recovers scene geometry from correspondences across sub-aperture views, yet its robustness to corruptions remains largely unexplored despite strong performance on clean benchmarks. Unlike single-image vision, LF corruptions can affect both individual-view appearance and the cross-view geometry carrying disparity, making visual severity alone insufficient to characterize robustness. We introduce LFDepth-C, a corruption-robustness benchmark built on the HCI 4D Light Field Dataset, covering 15 corruption types across five categories, two severity levels, and seven representative classical, supervised, and unsupervised methods. Beyond conventional depth metrics, we quantify corruption-induced changes in epipolar plane image (EPI) structure using a cross-view residual ratio and slope-recovery error. Results show that visual severity poorly predicts depth degradation: corruptions that preserve EPI geometry cause limited errors, whereas those that disrupt cross-view correspondences are substantially more harmful. Supervised methods are generally the most robust under zero-shot evaluation, while in-domain adaptation substantially improves several unsupervised methods, with gains varying across architectures and metrics. These findings identify cross-view geometry preservation as a key factor in LF depth robustness and provide a systematic basis for studying corruption-induced failures.