Photometric Stereo Using Sparse Bayesian Regression for General Diffuse Surfaces.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Conventional photometric stereo methods struggle with non-Lambertian surfaces and shadows.
• Existing techniques either degrade in diffuse regions or fail to account for outliers like shadows.

Purpose of the Study:

• To develop a robust, pixel-wise photometric stereo algorithm for non-Lambertian surfaces.
• To accurately estimate surface normals while separating diffuse and non-diffuse appearance components.

Main Methods:

• A novel method decomposes surface appearance into sparse non-diffuse and diffuse components.
• Piecewise linear approximation of the inverse diffuse model yields closed-form estimates.
• A hierarchical Bayesian model treats non-diffuse components as latent variables for accurate normal computation.

Main Results:

• The method stably and efficiently handles various non-Lambertian effects, including shadows and specularities.
• Accurate surface normal estimation is achieved by simultaneously separating diffuse and non-diffuse components.
• State-of-the-art performance demonstrated on both synthetic and real-world image datasets.

Conclusions:

• The proposed pixel-wise method offers a significant advancement in photometric stereo for complex surfaces.
• This approach provides a unified framework for handling non-Lambertian reflectance and image corruptions.
• The technique enables more accurate 3D surface reconstruction in challenging imaging conditions.