外観表現とイメージ操作による三次元物体認知

抄録

Two major problems in human object recognition are how we recognize objects from various viewpoints and how we memorize the shape of many objects. View-based object recognition theories have proposed that viewpoint independent recognition can be achieved by obtaining a certain number of views of an object. These theories do not involve the use of 3-dimensional information. In our previous research, however, we showed that the 3-dimensional structural information of objects could be utilized for recognition if enough time is available. The generalization from a familiar view to unknown views improved after trials only under the long reaction time conditions. According to the results, we supposed that there are two kinds of modules that compare the internal representation of objects and the input images. One is a 2-dimensional module that simply matches the images, and the other is a 3-dimensional module that involves transformation between relatively far viewpoints.
In this study, we first showed that 3-dimensional complexity of the objects affects the generalization range. The effect was seen only under the long reaction time conditions. This result strongly supports the above hypothesis. In the second part of this study, we replicated our previous result that the generalization range was broadened as a subject becomes more familiar with objects. And we found that the improvement mainly depends on familiarity with each category of objects rather than each object itself. These results cannot be explained by purely view-based theories which are modeled simply with the GRBF (Generalized Radial Basis Functions) network or its derivations, because in such theories each view of each object is independently acquired.