Automatic acquisition of exemplar-based representations for recognition from image sequences
Christian Wallraven, HH Bülthoff, Terence Sim · MPG.PuRe (Max Planck Society) · 2001
We present an exemplar-based object recognition system which is capable of on-line learning of representations of scenes and objects from image sequences. Local appearance features are used in a tracking framework to find `key-frames' of the input sequence during learning. The representation of the stored sequences which are used for recognition of novel images consists only of the appearance features in these key-frames and contains no further a-priori assumptions about the underlying sequences. The system is able to create sparse and extendable representations and shows good recognition performance in a variety of viewing conditions for databases of natural and synthetic image sequences.