Efficient Estimation of Multiple Illuminant Directions Using C-means Clustering and Self-correction for Augmented Reality

This paper presents a novel method to estimate multiple illuminant directions from a single image in augmented reality system. A square marker is used for 2D-3D registration, and a mirror sphere with known size is employed to detect light sources. Highlight pixels in captured sphere image are analyzed by using c-means clustering algorithm and its initialization with max-min distance method. The initialization estimates number of light sources and initial highlight areas centroids. c-means algorithm optimizes each position of the centroids. With the help of registration, multiple illuminant directions are calculated from the centroids and user’s viewpoint. Moreover, a self-correction course reduces estimation errors. Experimental results show that our approach is computationally efficient and multiple illuminant directions can be accurately obtained by it.