In this article we present a statistical framework for automatic classification and localization of 3D objects in 2D images. The new functionality of the framework allows us to use objects represented in different color spaces including gray level, RGB, and Lab formats. First, the objects are preprocessed and described by local wavelet features. Second, statistical modeling of these features under the assumption of their normal distribution is performed in a supervised way. The resulting probability density functions are determined by the maximum likelihood estimation. The density functions describe a particular object class from a particular training viewpoint. In the recognition phase, local feature vectors are computed from an image with an unknown object in an unknown pose. Those features are then evaluated against the trained density functions which yields the classes and the poses of objects found in the scene. Experiments performed for more than 40.000 images with real heterogeneous backgrounds have delivered very good classification and localization rates for all investigated object representations. Moreover, they brought us to interesting conclusions considering the general performance of statistical recognition systems for different image representations.