A Neurally-Inspired Approach to 3-D Visual Object Recognition

Abstract

A neurally-inspired 3-D visual object recognition system is described called SEEMORE, whose goal is to reliably identify common objects from a large known set — independent of viewing angle, distance, and partial occlusion. Three distinctive features of this view-based approach are: (1) use of a shallow but very high- dimensional feedforward network of simple filtering operations, (2) simultaneous use of cues from several different visual sub-modalities, i.e. color, shape, texture, etc., and (3) inclusion of objects that are rigid, non-rigid, articulated, and/or statistical in nature. Preliminary results have been obtained using combinations of color and shape-based features. In response to a testset of 150 novel views of 50 known objects presented individually in color video images, SEEMORE currently identifies the object correctly 83% of the time (chance is 2%). Six non-rigid objects, including a telphone cord, a bicycle chain, two scarves, a grape cluster, and a maple-leaf cluster were included in the training set; all 18 novel test views of these objects were correctly recognized. Recognition time on a Sparc-2 is a non-optimized 60 seconds including all image processing. Extension to the case of multiple object scenes is discussed.