PolarGS: Polarimetric Cues for Ambiguity-Free Gaussian Splatting With Accurate Geometry Recovery

Area of Science:

• Computer Vision
• Computer Graphics
• Optics

Background:

• 3D Gaussian Splatting (3DGS) achieves high geometric accuracy but struggles with reflective and textureless surfaces.
• Photometric ambiguities in these regions degrade reconstruction quality.
• Polarization of light offers complementary cues for surface orientation and resolving ambiguities.

Purpose of the Study:

• To introduce PolarGS, an optics-aware extension of 3DGS that incorporates polarization information.
• To enhance geometric accuracy in challenging surface reconstruction scenarios.
• To resolve photometric ambiguities using polarization priors.

Main Methods:

• Developed a polarization-guided photometric correction strategy using Degree of Linear Polarization (DoLP) to refine reflective Gaussians.
• Introduced a polarization-enhanced Gaussian densification mechanism integrating Angle and Degree of Linear Polarization (A&DoLP) for textureless area geometry recovery.
• Integrated a PatchMatch-based depth completion process for back-projection and fusion of new Gaussians.

Main Results:

• PolarGS significantly improves geometric accuracy in photometrically ambiguous regions.
• The method demonstrates superior performance compared to state-of-the-art 3DGS techniques.
• Achieved more complete and accurate 3D reconstructions, especially on reflective and textureless surfaces.

Conclusions:

• PolarGS effectively leverages polarization as an optical prior to overcome limitations of RGB-based 3DGS.
• The proposed modules enhance photometric consistency and geometry recovery.
• PolarGS offers a framework-agnostic solution for superior 3D surface reconstruction.