3D_DEN: Open-ended 3D Object Recognition using Dynamically Expandable\n Networks

Abstract

Service robots, in general, have to work independently and adapt to the\ndynamic changes happening in the environment in real-time. One important aspect\nin such scenarios is to continually learn to recognize newer object categories\nwhen they become available. This combines two main research problems namely\ncontinual learning and 3D object recognition. Most of the existing research\napproaches include the use of deep Convolutional Neural Networks (CNNs)\nfocusing on image datasets. A modified approach might be needed for continually\nlearning 3D object categories. A major concern in using CNNs is the problem of\ncatastrophic forgetting when a model tries to learn a new task. Despite various\nproposed solutions to mitigate this problem, there still exist some downsides\nof such solutions, e.g., computational complexity, especially when learning\nsubstantial number of tasks. These downsides can pose major problems in robotic\nscenarios where real-time response plays an essential role. Towards addressing\nthis challenge, we propose a new deep transfer learning approach based on a\ndynamic architectural method to make robots capable of open-ended learning\nabout new 3D object categories. Furthermore, we make sure that the mentioned\ndownsides are minimized to a great extent. Experimental results showed that the\nproposed model outperformed state-of-the-art approaches with regards to\naccuracy and also substantially minimizes computational overhead.\n