Polarimetric Multi-View Inverse Rendering.

Area of Science:

• Computer Vision
• Computational Imaging
• 3D Reconstruction

Background:

• Polarization imaging offers rich surface information.
• Angle of Polarization (AoP) and Degree of Polarization (DoP) relate to surface normals.
• Existing 3D reconstruction methods often lack detailed surface normal accuracy.

Purpose of the Study:

• To develop a novel 3D reconstruction method leveraging polarimetric information.
• To integrate geometric, photometric, and polarimetric cues for enhanced accuracy.
• To address ambiguities in polarization measurements for robust surface normal estimation.

Main Methods:

• Utilized Structure-from-Motion and Multi-View Stereo for initial 3D model and camera pose estimation.
• Proposed a novel polarimetric cost function to refine surface normals using AoP and DoP.
• Optimized photometric rendering and polarimetric errors on multi-view color-polarization images.
• Weighted polarimetric cost based on DoP reliability to handle azimuth ambiguities.

Main Results:

• Demonstrated accurate 3D shape reconstruction using the Polarimetric MVIR method.
• Achieved detailed surface normal estimation without prior material or lighting knowledge.
• Validated performance on both synthetic and real-world datasets.

Conclusions:

• Polarimetric MVIR effectively reconstructs detailed 3D shapes by integrating diverse image cues.
• The method shows robustness to ambiguities in polarization data.
• This approach advances 3D reconstruction capabilities for complex surfaces.