Color Normalization and Object Localization

L. Csink,D. Paulus,U. Ahlrichs,B. Heigl
Abstract:The distribution of color values in color images depends on the illumination which varies widely under real‐world conditions. Object recognition based on color information is expected to compensate for these effects. Several algorithms have been reported for color constancy or color normalization. We shortly review some of these: color rotation in various color spaces, whitening, and comprehensive color normalization. Then we use histogram backprojection to find our objects in various scenes using each of the above methods for normalization. Object positions are hand‐segmented in real‐world office scenes, in which illumiation varied as well as different cameras were used. These positions are compared to the automatically computed results, using each algorithm outlined above, or no normalization. In some cases depending on the object, color normalization enhanced finding rate. The normalization module is integrated into our complete system for knowledge based visual exploration of scenes.
What problem does this paper attempt to address?