Abstract
Service robots, in general, have to work independently and adapt to thedynamic changes in the environment. One important aspect in such scenarios isto continually learn to recognize new objects when they become available. Thiscombines two main research problems namely continual learning and 3D objectrecognition. Most of the existing research approaches include the use of deepConvolutional Neural Networks (CNNs) focusing on image datasets. A modifiedapproach might be needed for continually learning 3D objects. A major concernin using CNNs is the problem of catastrophic forgetting when a model tries tolearn new data. In spite of various recent proposed solutions to mitigate thisproblem, there still exist a few side-effects (such as time/computationalcomplexity) of such solutions. We propose a model capable of learning 3Dobjects in an open-ended fashion by employing deep transfer learning-basedapproach combined with dynamically expandable layers, which also makes surethat these side-effects are minimized to a great extent. We show that thismodel sets a new state-of-the-art standard not only with regards to accuracybut also for computational complexity.