Mixture of spherical distributions for single-view relighting

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Estimating scene illumination and object reflectance from images is crucial for realistic computer graphics and image manipulation.
• Previous methods often struggle with unknown numbers of light sources or complex reflectance properties.

Purpose of the Study:

• To develop a unified framework for simultaneously estimating illumination and reflectance properties from single-view images.
• To enable relighting of single-view images by accurately characterizing light sources and object surfaces.

Main Methods:

• Representing scene illumination as a mixture of von Mises-Fisher distributions using a novel spherical specular reflection model.
• Estimating mixture model parameters (number of light sources, surface roughness) and refining them using the Torrance-Sparrow reflection model.
• Utilizing directional statistics for a robust estimation framework.

Main Results:

• Simultaneous estimation of the number, intensities, and directions of multiple light sources.
• Accurate estimation of object specular reflectance properties (surface roughness).
• Successful relighting of single-view images by manipulating estimated light source parameters.

Conclusions:

• The proposed method offers a unified, statistically robust approach for illumination and reflectance estimation.
• This framework significantly advances the capabilities of single-image relighting and scene understanding.
• The method effectively handles an unknown number of point light sources and complex specularities.