Efficient Photometric Stereo Using Kernel Regression.

Area of Science:

• Computer Vision
• Computer Graphics
• Surface Reconstruction

Background:

• Photometric stereo estimates surface normals from multiple images with varying light directions.
• Current methods for non-Lambertian reflections are computationally intensive due to iterative or optimization frameworks.

Purpose of the Study:

• To propose an efficient photometric stereo method using kernel regression.
• To address the computational expense of existing methods for non-Lambertian surfaces.
• To achieve state-of-the-art accuracy in surface normal estimation.

Main Methods:

• Utilizes kernel regression, transformable to an eigendecomposition problem.
• Employs a variable kernel parameter per surface point for general reflectances.
• Automates kernel parameter selection using leave-one-out cross-validation.
• Accelerates leave-one-out cross-validation with fast matrix computation and normal initialization.

Main Results:

• The proposed method is computationally efficient.
• Achieves state-of-the-art accuracy in surface normal estimation.
• Validated on synthetic and real surfaces with diverse reflectances.

Conclusions:

• The kernel regression-based photometric stereo method offers a computationally efficient and accurate solution.
• The approach effectively handles general reflectances and achieves high accuracy.
• Demonstrates significant improvements over existing iterative and optimization-based methods.