Part of Advances in Neural Information Processing Systems 8 (NIPS 1995)
Bartlett Mel
A neurally-inspired visual object recognition system is described called SEEMORE, whose goal is to identify common objects from a large known set-independent of 3-D viewiag angle, distance, and non-rigid distortion. SEEMORE's database consists of 100 ob(cid:173) jects that are rigid (shovel), non-rigid (telephone cord), articu(cid:173) lated (book), statistical (shrubbery), and complex (photographs of scenes). Recognition results were obtained using a set of 102 color and shape feature channels within a simple feedforward network ar(cid:173) chitecture. In response to a test set of 600 novel test views (6 of each object) presented individually in color video images, SEEMORE identified the object correctly 97% of the time (chance is 1%) using a nearest neighbor classifier. Similar levels of performance were obtained for the subset of 15 non-rigid objects. Generalization be(cid:173) havior reveals emergence of striking natural category structure not explicit in the input feature dimensions.