PDN-Net: Multisensor Prior-Guided Joint Learning for Depth and Normal Estimation

Depth and surface normal estimation are fundamental to 3-D scene understanding, providing global structure and local geometric details. However, current methods often struggle with sensor limitations, leading to noisy or incomplete data, and fail to account for the geometric relationship between depth and normals, resulting in inconsistent predictions. To address these challenges, we propose PDN-Net, a multisensor prior-guided joint learning framework that extracts geometric priors from sensor depth, RGB, and polarization cues, integrating them through a joint learning paradigm to optimize both depth and normal estimation simultaneously. We introduce a multisensor prior extraction mechanism that leverages physics-informed modeling to capture sensor characteristics and cross-modal relationships, enabling reliable geometric priors through semantic-guided depth completion and polarization-based normal inference. Building on these priors, we design a joint learning paradigm that facilitates structured feature interactions and enforces geometric consistency through a progressive feature interaction framework and geometry-based refinement. Experiments on the HAMMER dataset demonstrate that PDN-Net achieves state-of-the-art results in depth completion and normal estimation, particularly in challenging scenarios involving reflective and transparent surfaces, demonstrating the effectiveness of multisensor priors and geometry-consistent joint learning.