In daily life humans easily distinguish large numbers of objects based on theirvisual appearance. Despite extensive efforts in recent years, modemrecognition systems are unable to robustly reproduce this capability. The mainproblem is that any object can have an infinite number of appearances, due todifferent viewing positions, lighting conditions and occlusion. A technicalsystem must leam to generalize over these irrelevant variances, whileconcentrating on the meaningful object information. This is done by the socalledfeature extraction.In this work I investigate two new methods for feature leaming that shouldovercome the limitations of existing approaches. The first method is based onthe holistic appearance of objects and tries to combine the advantages ofsupervised and unsupervised learning. I show for a constraint scenario that theobtained features improve the recognition performance. However the rigidnessof the holistic coding prevents the application of the method to more complex,real world scenarios. Because of this in the second approach I use a moreflexible representation that focuses on the presence of local object parts. Afterproposing a new supervised feature selection method, I show that the resultingrepresentation yields a strong performance on various object databases andavoids some drawbacks of established recognition approaches. Finally Iintegrate the approach into areal-time recognition system that is the first one torobustly identify about 120 objects of arbitrary shape and texture under 3Drotation in front of cluttered background, and thus marks a major step towardsinvariant object recognition.