View Planning For Cloud-Based Active Object Recognition

Abstract

One of the central problems in computer vision and robotics is to recognize objects in a scene. State-of-the-art algorithms for object recognition are extremely data intensive. Cloud technologies hold the promise to make such algorithms available to robots with limited computation capabilities. However, collecting and transferring large amounts of data with such robots remains a challenge. In this work, we investigate the possibility of enabling cloud based object recognition by carefully planning the robot's viewpoints. While view planning techniques for object recognition exist, such techniques are too costly to be executed by robots with limited capabilities which are the robots which would benefit most from cloud-based techniques.

In this paper, we present evidence for the existence of universal viewpoints: a small number of viewpoints which guarantee accurate object recognition regardless of the object pose. Our experiments with real data provide evidence that such view points exist for common objects. Hence, view-planning for object recognition can be performed in an open-loop fashion without the need for running costly algorithms on small robots.