Abstract:

Light field cameras capture synchronized spatial-angular information, offering a new paradigm for 3D visual perception. Depth estimation, as a fundamental task in light field analysis, underpins critical applications including 3D reconstruction and visual odometry. However, occlusion-induced estimation errors remain a persistent challenge. This paper proposes an occlusion-aware depth estimation framework featuring two novel modules: entropy-based occlusion mask pre-computation method and a viewpoint screening-driven depth estimation algorithm. First, we model light field occlusion by analyzing the entropy of micro-images array, constructing a local entropy extremum-based occlusion mask pre-computation approach. This method addresses the limitations of conventional techniques in characterizing occluded regions. Subsequently, viewpoint screening is employed to eliminate interference from occluded perspectives, effectively reducing estimation errors and decreasing the proportion of disparity outliers exceeding 0.03 pixels compared to other methods. The core contribution lies in establishing a theoretical connection between entropy and light field occlusion, enabling a robust occlusion-aware depth estimation framework. Experiments on 4D light field benchmark demonstrate that our method achieves superior performance in mean absolute error (MAE) and 25th error percentile (Q25). Comparative ablation studies confirm the efficacy of the entropy-driven occlusion mask, highlighting the critical role of entropy in advancing light field occlusion.

Key words: light field, depth estimation, entropy