A Feature-based Aspect Recognition Algorithm
Di Wang, Andrew H. Fagg · 2007
Robotic vision is an important domain in robotics research that can be used to facilitate solutions to many other robotic problems, such as grasping and manipulation [2]. In this presentation, we intend to develop an algorithm that can be used to estimate both the identity and the pose of an object so that a robot can reach out and grasp it. The challenge is how to construct visual models of objects that support pose estimation given experience with manipulating the objects. We propose a method of visually identifying the direction from which the object is viewed [3] (which is referred to as pose recognition or aspect recognition). The aspect recognition model is trained using experience from simultaneously manipulating and watching the target object. Constellations of image features are used to represent the appearance of the object at certain viewing angles. Clusters of viewing angles in which a given constellation is viewable are represented as probability density functions defined over the unit sphere [1]. By combining the evidence from several observed constellations, one can infer an estimate of viewing angle given a novel image. We demonstrate empirically that constellations which are locally robust and yet selective over all possible viewing angles lead to improved aspect recognition rates over the ones that do not satisfy these properties.