In recent times, Convolutional Neural Network (CNN) based approaches have performed exceptionally well in many computer vision related tasks, including classification and segmentation. These approaches have shown that given enough training data and time, they can often perform at a level significantly higher than the alternative methods. However, in the context of robotic learning, it is commonly the case that both time and training data are limited. In this work, we propose a learning approach that is more suitable for robotic learning; it substantially reduces the time required to learn and provides much higher performance when training data is limited. Our method combines random forests with deep convolution networks, leveraging the strengths of both frameworks. We develop a method for generating derivatives from our highly non-linear forest classifier which in turn enables training of the CNN. Furthermore, our method allows leaf distributions in the ensemble classifier to be trained jointly with one another using Stochastic Gradient Descent (SGD), allowing for parallel training of a large number of tree classifiers at once. This results in a drastic increase in training speed. Our model demonstrates significant performance improvements over pure deep learning methods, notably on datasets with limited training data. We apply our method to the outdoor and indoor segmentation datasets of KITTI and NYUv2-40, outperforming multiple pure deep learning methods whilst using a fraction of training time normally required.
展开▼
