Stephan Hasler, "Learning Features for Robust Object Recognition", 2010.

Humans can easily recognize a very large number of previously seen objects. Despite extensive efforts in recent years, the principles underlying this capability are hardly understood and modern object recognition systems are far from reaching human performance. The main reason for this is that the visual appearance of an object is strongly influenced by various conditions. So depending on the viewing angle different parts of an object are visible or occluded. The resolution of details is influenced by the viewing distance and the lighting conditions, where the latter ones also determine the perceived color of an object and can cause artifacts, like shadows and reflections. Also there is usually more than one object visible at a time. These objects may overlap each other in the image plain, which also leads to occlusion or to other distortions in case of transparent objects. Together all these possible variations make it impossible to capture the characteristic of an object by means of a single snapshot. Instead, a recognition system has to implement a dedicated process to extract the information that is specific for an object and thus can be used to separate it from other ones.