Physics-Aware Baseline Decoupling for Light Field Depth Estimation

Light field (LF) depth estimation recovers scene geometry from dense 4D spatial-angular observations and supports applications such as 3D reconstruction and refocusing. However, its performance is fundamentally constrained by the baseline-dependent nature of LF imaging. Specifically, medium-baseline views offer more stable correspondences due to smaller disparity variations, which reduces matching ambiguity for global geometry. In contrast, large-baseline views introduce stronger disparity that enhances detail discrimination, but also amplify occlusions and non-Lambertian effects. This trade-off poses a fundamental challenge: how to effectively utilize different baseline views without introducing conflicting cues. Existing methods often treat all sub-aperture images uniformly, ignoring baseline-specific characteristics, which leads to feature conflicts and degraded depth estimates. To address this, we propose Physics-Aware Baseline Decoupling, a principled framework that disentangles and leverages view subsets based on their physical properties. We first perform baseline-consistent geometry initialization using polar positional encoding on medium-baseline views to estimate a reliable coarse depth. This serves as a geometric prior for aligning large-baseline outer views, where we compute a perpixel risk score to selectively retain the topK low-risk views. These views are fused with reliability-aware weights, followed by a residual refinement in a narrow-band cost volume. This decoupling strategy enables fine-scale detail recovery only where physical consistency is ensured. We implement this framework as PANet, which achieves state-of-the-art results on multiple synthetic and real-world LF datasets, demonstrating superior depth accuracy and robustness.