Buluswar [3] showed that pixel classification based on samples of a target object under varying conditions can be an effective way of color recognition without using surface reflectance or illumination models. Techniques such as multivariate decision trees and neural networks were used for classifying pixels in indoor and outdoor images. While the method was shown to be successful in a number applications, the primary problem with it is that training samples of both the target object and miscellaneous background objects need to be representative of the gamut of the respective apparent colors.