Open-Ended Fine-Grained 3D Object Categorization by Combining Shape and Texture Features in Multiple Colorspaces

As a consequence of an ever-increasing number of service robots, there is a\ngrowing demand for highly accurate real-time 3D object recognition. Considering\nthe expansion of robot applications in more complex and dynamic environments,it\nis evident that it is not possible to pre-program all object categories and\nanticipate all exceptions in advance. Therefore, robots should have the\nfunctionality to learn about new object categories in an open-ended fashion\nwhile working in the environment.Towards this goal, we propose a deep transfer\nlearning approach to generate a scale- and pose-invariant object representation\nby considering shape and texture information in multiple colorspaces. The\nobtained global object representation is then fed to an instance-based object\ncategory learning and recognition,where a non-expert human user exists in the\nlearning loop and can interactively guide the process of experience acquisition\nby teaching new object categories, or by correcting insufficient or erroneous\ncategories. In this work, shape information encodes the common patterns of all\ncategories, while texture information is used to describes the appearance of\neach instance in detail.Multiple color space combinations and network\narchitectures are evaluated to find the most descriptive system. Experimental\nresults showed that the proposed network architecture out-performed the\nselected state-of-the-art approaches in terms of object classification accuracy\nand scalability. Furthermore, we performed a real robot experiment in the\ncontext of serve-a-beer scenario to show the real-time performance of the\nproposed approach.\n