Combined Deep and Active Learning for Online 3D Object Recognition

Abstract

Deep learning methods have received lots of attention in research on 3D object recognition. Due to the lack of training data, many researchers use pre-trained Convolutional Neural Networks (CNNs) and either extract the output of one of the last layers as features or fine-tune the networks on their data. We achieve superior results with a method that fine-tunes a CNN before feature extraction for RGB data. Combined with extracted features from depth data and reducing the features’ dimensionalities, we improve the state-of-the-art accuracy on the University of Washington RGB-D Object dataset [Lai+11], using a support vector machine (SVM). Furthermore, we evaluate the impact of different learning rates (LRs) when fine-tuning a CNN. Our results show that the selection of a suitable LR is crucial to the success of a network. Instead of SVM as a classifier, we also use the Mondrian forest (MF), an online classifier, which can be updated over time as soon as more data is available.