Title: Leveraging Structure-Aware Consistency Priors for Shape from Polarization in Complex Environments

Abstract:

Estimating surface normals from single-view polarization imagery within complex media presents a significant challenge. This study utilizes ice as a primary example of such an environment, where complex light-matter interactions result in a nonlinear relationship between polarization data and surface normal vectors. To overcome this difficulty, we introduce a structure-aware polarization prior grounded in autocorrelation functions, which effectively models the local spatial consistency of the Angle of Linear Polarization (AoLP). Furthermore, we propose IceSfP, a dual-branch architecture that synthesizes raw polarization inputs with these priors through multi-scale feature fusion and cross-modal attention mechanisms, thereby facilitating precise surface normal recovery in complex media. For validation purposes, we have compiled the first real-world dataset for ice-based Shape from Polarization (SfP). Our experiments demonstrate that the proposed approach surpasses current state-of-the-art methods across every evaluation metric, securing a Mean Absolute Error (MAE) of 16.01 degrees—a substantial improvement of 2.74 degrees over the runner-up technique. This framework offers a broadly applicable solution for achieving high-precision geometric perception in challenging optical environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC