High-precision polarization-based 3d reconstruction based on fine depth map

In complex scenarios, traditional vision-based 3D reconstruction techniques often fail to yield satisfactory results for low-texture and high-reflectivity objects. Using polarization information to reconstruct such objects is convenient and effective, yet sole reliance on it brings challenges like surface normal vector ambiguity and difficulties in normal vector integration. To address these issues, this paper proposes a high-precision 3D reconstruction method integrating polarization information and binocular vision. First, in the binocular system, a single-shot defocused plane camera extracts the target surface’s polarization degree and angle; combined with the Fresnel reflection model, an initial ambiguous normal vector (reflecting local normal probability distribution but with directional ambiguity) is calculated. Polarization images of the same scene from the system are rectified, and disparity information is obtained via a stereo matching network—its cascaded feature network and local attention search mechanism enable more accurate pixel-wise depth construction for fine depth data. Next, a geometric constraint-based guiding normal vector is built to clarify surface normal spatial orientation via binocular disparity constraints. Additionally, fine depth is used as prior information to resolve polarization normal vector ambiguity. Finally, a normal vector fusion algorithm is designed: it uses the guiding normal vector’s spatial geometric constraints to eliminate ambiguous normal vector directional ambiguity, enabling accurate surface normal calculation (avoiding normal vector integration) and constructing a high-precision 3D reconstruction model via dense point cloud matching. Experimental results show the proposed algorithm improves polarization imaging quality, effectively restores lost depth details, smooths low-texture and high-reflectivity areas, and enhances 3D reconstruction quality.