Comparison of active SIFT-based 3D object recognition algorithms

Abstract

Active object recognition aims to manipulate the sensor and its parameters, and interact with the environment and/or the object of interest in order to gather more information to complete the 3D object recognition task as quickly and accurately as possible. It can leverage the mobility of robotic platforms to capture additional viewpoints about an object as single images are not always sufficient especially if objects appear in cluttered human environments. Active vision algorithms should reduce the number of viewpoints required to recognise an object and hence reduce the computational time as well. This paper compares two active object recognition systems. Both systems use SIFT features for object recognition, but use contrasting models, update and viewpoint selection strategies. The methods for integrating information across views used by the two systems are investigated. This is essential as this module is used to select the next best viewpoint. The number of viewpoints and the time taken to recognise objects are used to compare the performance of these two methods.