Uncalibrated, Two Source Photo-Polarimetric Stereo

Area of Science:

• Computer Vision
• Photometry
• Surface Reconstruction

Background:

• Accurate 3D shape and material property estimation is crucial for computer vision applications.
• Existing methods often require controlled, calibrated lighting conditions, limiting their real-world applicability.
• Uncalibrated shape from shading and polarization (photo-polarimetric) methods face challenges with unknown illumination variations.

Purpose of the Study:

• To develop robust methods for photo-polarimetric shape estimation under uncalibrated, varying illumination.
• To introduce novel constraints and a unified framework for solving shape estimation problems.
• To enhance the estimation of illumination, albedo, and refractive index alongside shape.

Main Methods:

• Formulating photo-polarimetric constraints based on surface partial derivatives within a unified partial differential equation system.
• Developing a linear least squares approach by eliminating non-linearities for discrete problem solving.
• Implementing a minimal, combinatorial approach with RANSAC for robust two-source illumination estimation (light direction and intensity).
• Introducing a new method for polarization image estimation from multichannel data.
• Providing methods for albedo and refractive index estimation.

Main Results:

• Demonstrated a unified system of partial differential equations encompassing previous work as a special case.
• Successfully solved discrete shape estimation problems using linear least squares after non-linearity elimination.
• Achieved robust estimation of light direction and intensity using a combinatorial approach with RANSAC.
• Validated the effectiveness of proposed methods for estimating polarization images, albedo, and refractive index.
• Showcased improved performance over state-of-the-art methods on both synthetic and real-world datasets.

Conclusions:

• The proposed photo-polarimetric methods provide a significant advancement for shape and material property estimation in uncalibrated environments.
• The unified framework and novel constraints offer a more comprehensive and accurate approach to 3D reconstruction.
• The developed techniques demonstrate practical utility and improved performance for real-world computer vision challenges.