New PhD Vacancy: Implicit Shape Representations in Computer Vision
News
University of Bath Department of Computer Science
Funded PhD Project (UK & International Students)
About the Project
The University of Bath is inviting applications for a PhD project on the use of implicit shape representations in computer vision in the Department of Computer Science commencing 30 September 2024.
The successful student will be part of the Centre for the Analysis of Motion, Entertainment Research and Applications (CAMERA) which performs world-leading multi-disciplinary research in Intelligent Visual and Interactive Technology. Funded by the EPSRC and the University of Bath, CAMERA exists to accelerate the impact of fundamental research being undertaken at the University in the Departments of Computer Science, Health and Psychology. The successful candidate will work closely work with the experts from CAMERA and potentially with collaborators from the University of Bristol and project partners associated with the MyWorld programme. The ambititious MyWorld project is funded by the UKRI Strength in Places fund bringing together 30 partners from Bristol and Bath’s creative technologies sector and world-leading academic institutions to create a unique cross-sector consortium.
Overview of this Project:
Finding an appropriate digital representation for real-world objects has challenged many researchers and a variety of approaches have been presented over the years. Traditionally shapes have been represented explicitly using voxels, point clouds or meshes. Although highly efficient methods have been developed for inference, they suffer from discretisation and lack of mathematical description away from the surface.
The last few years have seen Implicit Neural Representations (INR) gain popularity, due to the fact they model surfaces using a continuously differentiable function, which can be used to produce high-resolution outputs. In particular, Neural Radiance Fields (NeRFs) have revolutionized novel-view synthesis captured with multiple photos or videos, producing photo-realistic, high-resolution, and view-consistent scenes. While Signed Distance Fields (SDFs) have proved extremely effective at encoding efficient representations of surface geometry.
Due to the fully differentiable nature of these representations, it has been demonstrated that these representations can be exploited for a variety of tasks of downstream tasks. A particularly popular usage is as part of a generative model synthesising high-quality human shape, pose and dynamics.
Despite significant progress towards improving training times and memory efficiency, the sampling required for rendering is costly and can result in noise. Recently, Gaussian Splatting has demonstrated the capacity to maintain all the favourable characteristics of volumetric radiance fields, while eliminating the neural network facilitates efficient training and real-time rendering. This has opened a series of exciting potential new avenues for exploration.
What are you going to do?
We seek a PhD Candidate that will contribute to research at the intersection of implicit shape representations and Machine Learning (ML) pipelines for a range of downstream tasks.
While this being an incredibly fast-moving field, there still remain a plethora of potential novel use cases. For example, despite these representations being easy to fit, current methods do not offer a way to incorporate the uncertainty of their reconstruction, which is a highly desirable property for real world tasks involving medical imaging and autonomous driving.
Modelling dynamic scenes is still at a very immature stage with significant scope for innovation. This is especially the case when incorporating challenging real-world data which contains noisy measurements, for example, video taken in low light or noisy point clouds from range scanners.
A frontier with huge potential is to further bridge recent successes in differentiable surface modelling and rendering with other ML pipelines in order to investigate novel solutions for a range of common computer vision, geometric modelling and animation tasks.
For further details, please see https://www.ndfcampbell.org/opportunities/
