Item

Abstract

One major difficulty in computational object recognition lies in the fact that a 3D object can be seen from an infinite number of viewpoints. Thus, the issue arises that objects with different 3D shapes often share similar 2D views. Humans are able to resolve this kind of ambiguity by
producing additional views through object manipulation or self movement. In both cases the action made provides proprioceptive information linking the visual information retrieved from the obtained views. Following this process, we combine visual and proprioceptive information to increase recognition performance of a computer vision system. In our approach we place a 3D model of an unknown object in the hand of a simulated anthropomorphic robot arm. The robot now executes a predefined exploratory movement to acquire a variety of different object views. To assure computational tractability, a subset of representative views is selected using the Keyframe concept by Wallraven et al. (2007). Each remaining frame is then annotated with the respective proprioceptive configuration of the robot arm and the transitions between these configurations are treated as links between object views. For recognizing objects this representation can be used to control the robot arm based on learned data. If both proprioceptive and visual data agree on a candidate, the
object was recognized successfully. We investigated recognition performance using this method. The results show that the number of misclassified results decreases significantly as both sources â visual and proprioceptive â are available, thus demonstrating the importance of a combined space of visual and proprioceptive information.