PhD Studentship: Seeing the Light: Computer Vision in the Real World (FINLAYSON_U17SCI50)
University of East Anglia - School of Computing Sciences
|Funding for:||UK Students, EU Students|
|Placed on:||10th November 2016|
|Closes:||10th May 2017|
Start Date: October 2017
Supervisor: Prof Graham Finlayson
Project description: We’d like to see (to recognise, navigate and track) the objects in the world as being independent of the illumination environment. Yet, light and surface play equal and symmetrical roles in image formation. Physically, a white piece of paper imaged under blue light is identical to a blue paper viewed under white light [i]. Perhaps unsurprisingly, this ambiguity makes it hard to divine that we are seeing the same object when it is seen inside and outside of shadow (the illumination colour makes the image RGBs very different). After many decades of research a general solution to this multiple-light/multiple-surface disambiguation problem has not been found.
This said, good progress has been made to solving the simplified problem, where there is a single simple light (same colour everywhere) illuminating a scene containing many surfaces [ii]. This project begins by investigating how this simple algorithm might be repurposed to work in more general (and typical) lighting conditions. From a physics point of view, the recent recasting of the image formation problem using the concept of homographies [iii] (previously only thought relevant to the geometry of computer vision) provides a framework for investigating the general light and surface disambiguation problem. Any developed theory might then be implemented in a variety of ways ranging from simple physics-based algorithms to statistical algorithms (plausibly, such as those using deep learning [iv]).
While the algorithms to be developed in this research might be widely applicable in computer vision (e.g. for recognition, navigation and tracking) the key application considered in this project will be digital photography. Simply, we seek to manipulate the illumination field in images to produce more pleasing photos. This work runs in collaboration with Apple Inc.
Person specification: Minimum 2:1 in Computer Science, Physics, Engineering, Psychology or other numerate discipline (assuming basic competency in programming)
Funding notes: This PhD studentship is jointly funded for three years by Faculty of Science and Apple. Funding comprises home/EU fees, an annual stipend of £14,296 and £1000 per annum to support research training. Overseas applicants may apply but are required to fund the difference between home/EU and overseas tuition fees (in 2016/17 the difference is £9,679 for the School of Computing Sciences but fees are subject to an annual increase).
Share this PhD
Type / Role:
South East England