Case studies: Machine Learning

Semantic segmentation of 2D images

by Shuai Zheng (Philip Torr group, www.robots.ox.ac.uk/~tvg)

With an application to augmented-reality glasses to help the partially sighted, this study tested new segmentation algorithms for 2D images built on top of the Caffe deep learning library.  The study benefited from the availability of compute nodes with Nvidia K40 and K80 GPUs as part of the ARC service.

Semantic segmentation of 2D images using new algorithms build with the Caffe deep learning library.

Further information can be found at Shuai's robots website and at inverse.com.

Monocular vision odometry

by Sen Wang (Paul Newman’s group, robots.ox.ac.uk/~pnewman)

This novel work in visual odometry aimed to achieve end-to-end pose estimation from monocular vision using Deep Learning.  Visual odometry -- as one of the essential techniques for pose estimation, robot localisation and navigation -- has attracted significant interest in both the computer vision and robotics communities over the past few decades.  This study has shown competitive performance compared with state-of-the-art methods and it is expected to be a viable technique towards improving the performance of visual odometry in terms of accuracy and robustness.

The use of ARC was essential in the availability of GPU-based resources for all computational platforms (Theano, Keras and Caffe) as well as for providing temporary fast storage for the large image datasets used in training and testing deep neural network models.  Sen says that "ARC not only significantly accelerates training of deep neural networks, but also enables us to develop more sophisticated neural networks to achieve better results."

A schematic representation of the computational workflow on the ARC service. The neural network takes a sequence of raw RGB images as input and directly estimates their poses (position and orientation) in an end-to-end manner.

Comparison between the benchmark odometry calculations using the deep learning based algorithm and other established methods.

Benchmark sequence of 2D monocular images.