Color constancy using 3D scene geometry derived from a single image.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Color constancy is crucial for accurate image analysis but is an ill-posed problem.
• Existing algorithms rely on simplifying assumptions like gray-world or white-patch, limiting their effectiveness.
• A robust color constancy method needs to account for varying illumination conditions within an image.

Purpose of the Study:

• To develop a color constancy algorithm that adapts to varying illumination by leveraging 3D image geometry.
• To classify images into distinct geometric stages for tailored color constancy application.
• To combine multiple color constancy algorithms based on depth and local image statistics.

Main Methods:

• Utilized 3D geometry models to classify images into rough geometric stages.
• Employed hard and soft segmentation to divide images into stage regions based on depth.
• Investigated the relationship between depth, local image statistics, and color constancy to select appropriate algorithms per region.
• Proposed a method to combine color constancy algorithms based on image depth and statistics.

Main Results:

• The proposed method significantly outperforms state-of-the-art single color constancy algorithms, achieving nearly 50% improvement in median angular error.
• With a perfect classifier, the performance improvement reached 52% median angular error compared to the best single algorithm.
• The approach enables the estimation of multiple illuminations by differentiating between nearby and distant light sources.

Conclusions:

• Leveraging 3D geometry and image statistics provides a more robust approach to color constancy.
• Adaptive selection of color constancy algorithms based on image depth enhances performance.
• This method offers a promising direction for handling complex lighting scenarios and estimating multiple illuminations.