RGB-D Object Recognition from Hand-Held Object Teaching

Abstract

For RGB-D object recognition, conventional methods only focus on classification, which neglects the importance of humans for object segmentation and object concept learning in the interaction and has limitations when transferring the learned knowledge to general indoor scenes. In this paper, we propose a system that humans can teach robots object concepts and instances by the way of handheld object recognition, a newly research topic in RGB-D object recognition, then these concepts and instances are transferred to more general indoor scenes to recognize target objects. Unlike traditional approaches of RGB-D object recognition, human-machine interaction is considered in our system, in which robots can obtain the object region from object segmentation and enhance its own interactive learning ability. In addition, we propose an RGB-D object dataset to match the hand-held object dataset. In experiments, we use CNN to learn feature representation of RGB-D images in the hand-held training object dataset and RGB-D test object dataset respectively. Experimental results show that our system has a strong capability in RGB-D object recognition.