Adaptive Object Acquisition.
Gabriele Peters, Claus-Peter Alberts, Markus Briese, Sebastian Entian, Christian Gabriel, Zhiqiang Gao, Alexander Klandt, Thomas Leopold, Jan Schultze, Jeremias Spiegel, Jürgen Thyen, Martina Vaupel, Peter Voß, Qing Zhu · 2005
We propose an active vision system for object acquisition. The core of our approach is a reinforcement learning module which learns a strategy to scan an object. The agent moves a virtual camera around an object and is able to adapt dynam-ically to different conditions of its environment such as different objects and different purposes of the data acquisition by means of a reinforcement signal which rewards a chosen action with respect to the intendet purpose. The purpose of the acquisition we consider here is the reconstruction of non-acquired views. The learned scan path allows the generation of a sparse, view-based object representation which consists of some key views of the scan path. We present preliminary results from a project conducted with undergraduate students and show that the scan pattern obtained with the proposed method allows a better reconstruction of unfamiliar views than random scan paths. As the reward signal is based on local information at the current position of the camera this approach is an example for an organic-computing system with the emergence of a global strategy from local rules. 1 The Problem Computer vision, as well as computer graphics, deals with the visual appearance of objects. Among the topics of computer graphics is the generation of 3d models from real