Active humanoid vision and object classification

In this paper we study object learning and recognition on a humanoid robot with foveated vision. The developed approach is view-based and can learn viewpoint-independent representations for object recognition. The training data is collected statistically and in an interactive way where a human instructor freely shows the object from a number of different viewpoints. The proposed system was fully implemented and runs in real-time, which is essential for meaningful interaction with a humanoid robot.