Concurrent 3D Object Category Learning and Recognition Based on Topic Modelling and Human Feedback

摘要

In open-ended domains, autonomous robots must have the ability to continuously process visual information, and execute learning and recognition in a concurrent and interleaved fashion. Because the set of categories to be learned is not predefined, it is not feasible to assume that one can pre-program all object categories required by service robots. Topic modelling approaches usually construct the topics from a training set to recognize objects. However, in open-ended domains, the data available for training increases continuously. If limited training data is used, this might lead to non-discriminative topics and, as a consequence, to poor object recognition performance. This paper proposes an object recognition system capable of learning object categories as well as the topics used to encode objects concurrently and in an open-ended manner. This system provides a robot with the capabilities to, (i) use unsupervised object exploration to construct a dictionary of visual words for representing objects and (ii) conceptualize object experiences and learn new object categories using topic modelling and human feedback. To examine the performance of the system, an on-line evaluation protocol is used to assess the performance of the system in an open-ended setting. The experimental results show the fulfilling performance of this approach on different types of objects.