Shape from polarization via a physical prior-based deep fusion network with ambiguous surface normals

Area of Science:

• Computer Vision
• Optical Imaging
• 3D Reconstruction

Background:

• Shape from polarization (SfP) imaging is a precise 3D reconstruction technique.
• Current deep learning SfP methods struggle with complex scenes due to insufficient fusion of physical prior information.
• Degraded reconstruction quality is observed in challenging environments.

Purpose of the Study:

• To improve the accuracy and robustness of shape from polarization methods.
• To effectively integrate polarization information with ambiguous surface normals and specular confidence.
• To enhance 3D reconstruction quality in complex scenes and under adverse lighting.

Main Methods:

• A novel method integrating polarization, ambiguous surface normals, and specular confidence information.
• Calculation of ambiguous surface normals and specular confidence using a physical model.
• Development of a dual-branch deep fusion network for feature extraction and fusion.

Main Results:

• Significantly improved reconstruction accuracy in complex scenes.
• Accurate surface normal reconstruction under complex lighting and low-texture conditions.
• Enhanced robustness of shape from polarization methods.

Conclusions:

• The proposed method effectively fuses polarization and physical prior information for superior 3D reconstruction.
• The dual-branch deep fusion network achieves high accuracy and robustness.
• The method holds significant potential for applications in intelligent manufacturing, defect detection, and medical imaging.