In daily life humans easily distinguish large numbers of objects based on their visual appearance. Despite extensive efforts in recent years, modem recognition systems are unable to robustly reproduce this capability. The main problem is that any object can have an infinite number of appearances, due to different viewing positions, lighting conditions and occlusion. A technical system must leam to generalize over these irrelevant variances, while concentrating on the meaningful object information. This is done by the socalled feature extraction. In this work I investigate two new methods for feature leaming that should overcome the limitations of existing approaches. The first method is based on the holistic appearance of objects and tries to combine the advantages of supervised and unsupervised learning. I show for a constraint scenario that the obtained features improve the recognition performance. However the rigidness of 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 more flexible representation that focuses on the presence of local object parts. After proposing a new supervised feature selection method, I show that the resulting representation yields a strong performance on various object databases and avoids some drawbacks of established recognition approaches. Finally I integrate the approach into areal-time recognition system that is the first one to robustly identify about 120 objects of arbitrary shape and texture under 3D rotation in front of cluttered background, and thus marks a major step towards invariant object recognition.weiterlesen