Single Image 3-D Geometry Estimation using Unsupervised Machine Learning (GONGH1U19SF)

Most current deep Neural Networks are trained by supervised learning which usually requires massive amounts of manually labelled data (e.g. [i]). Unsupervised learning is a branch of machine learning that learns from unlabelled data without any involvement of human annotators. A toddler has never been taught how to perceive the 3-D world, yet a toddler’s ability of 3-D perception improves over time. It is self-reasoning, and this is what a true intelligence is about for the future Artificial Intelligence (AI) industry.

In this PhD project, we plan to develop an unsupervised machine learning framework to train a deep neural network for 3-D geometry (or depth image) reconstruction from a single image. The ground-breaking novelty of this project is that we do not require a pre-training stage or any labelled ground-truth data (such as the depth images captured by a 3-D scanner). Similar to a toddler’s self-learning, the only training data we use will be the common videos captured by an RGB camera, which can be easily collected from the Internet. We hope to train a machine 3-D geometry perception just by “observing the world”. We plan to achieve this by taking advantage of the between-frame cues such as geometry and illumination consistency [ii-iv]. We also investigate the possibility of using an image-to-image Generative Adversarial Network (GAN) [v]. The developed approach will be evaluated based on some public geometry reconstruction datasets (esp. for data captured from on-car cameras).

The outcome of this project is expected to provide powerful applications in many hot areas of AI such as self-driving cars/drones, mixed-reality engines (for Oculus, Hololens, and Magic Leap), and film visual effects.

A start date prior to October 2019 is possible but should be discussed with Dr Gong in the first instance.

For more information on the supervisor for this project, please go here: https://people.uea.ac.uk/en/persons/h-gong
The type of programme: PhD   
The start date of the project: October 2019
The mode of study: full-time
Entry requirements: Acceptable first degree in Computer Science, Mathematics, Physics (or related degree), the standard minimum entry requirement is 2:1