3D Diffusion Models for Generating and Understanding 3D Scenes. Diffusion models, such as DALL-E2 and Imagen, have achieved remarkable success in generating photorealistic images and hold promise to solve long-standing computer vision problems. However, 3D scene generation remains unexplored. This research project aims to bridge the gap by developing 3D diffusion models capable of generating complete 3D scenes. This will advance our theoretical understanding of diffusion in complex 3D environmen ....3D Diffusion Models for Generating and Understanding 3D Scenes. Diffusion models, such as DALL-E2 and Imagen, have achieved remarkable success in generating photorealistic images and hold promise to solve long-standing computer vision problems. However, 3D scene generation remains unexplored. This research project aims to bridge the gap by developing 3D diffusion models capable of generating complete 3D scenes. This will advance our theoretical understanding of diffusion in complex 3D environments and open up new possibilities for applications in fields such as virtual reality, architecture, and city planning. The proposed 3D diffusion models will also enhance the accuracy of computer vision tasks related to 3D scene understanding, such as object detection, tracking, and semantic segmentation.Read moreRead less
Shape4D: Modelling the Spatiotemporal Deformation Patterns in 3D Shapes. This research will develop new mathematical methods and algorithms that will enable the use of population-level longitudinal studies to model the spatial and temporal deformation patterns in 3D biological objects. Using novel geometric and deep learning techniques, it will create new methods that will allow the characterization of how the 3D shape of objects deforms with ageing, disease progression and interaction with thei ....Shape4D: Modelling the Spatiotemporal Deformation Patterns in 3D Shapes. This research will develop new mathematical methods and algorithms that will enable the use of population-level longitudinal studies to model the spatial and temporal deformation patterns in 3D biological objects. Using novel geometric and deep learning techniques, it will create new methods that will allow the characterization of how the 3D shape of objects deforms with ageing, disease progression and interaction with their environment, and the simulation of spatiotemporal deformations in anatomical organs. Benefits include a better understanding of growth processes, predictive models of how degenerative diseases progress and a computational framework that will assist in designing proper mitigation and intervention strategies.Read moreRead less
Tensor and Hypergraph Methods in Fitting Visual Data. This proposal will put an important class of clustering (extracting data that should fit a geometric model) on a more solid theoretical foundation. This will lead to better understanding of how to certify outcomes, efficiency, reliability etc. The type of clustering under consideration is relevant to many problems in machine learning and computer vision, as well as data mining and a wide variety of other settings.
New Paradigms for Robust Fitting: Kernelisation and Polyhedral Search. Outliers inevitably exist in visual data due to imperfect data acquisition or preprocessing. To enable computer vision applications that can perform reliably, robust fitting algorithms are necessary to counter the biasing influence of outliers. However, current robust algorithms are unsatisfactory: they are unreliable (due to using randomisation) or too computationally costly (due to using exhaustive search). This project wil ....New Paradigms for Robust Fitting: Kernelisation and Polyhedral Search. Outliers inevitably exist in visual data due to imperfect data acquisition or preprocessing. To enable computer vision applications that can perform reliably, robust fitting algorithms are necessary to counter the biasing influence of outliers. However, current robust algorithms are unsatisfactory: they are unreliable (due to using randomisation) or too computationally costly (due to using exhaustive search). This project will develop new robust algorithms to mitigate these shortcomings. It will do so by investigating two new paradigms of kernelisation and polyhedral search, which offer unprecedented theoretical insights into the problem. The outcomes will contribute towards computer vision applications that are more practical and reliable.Read moreRead less
Deep visual understanding: learning to see in an unruly world. Deep Learning has achieved incredible success at an astonishing variety of Computer Vision tasks recently. This project aims to convey this success into the challenging domain of high-level image-based reasoning. It will extend deep learning to achieve flexible semantic reasoning about the content of images based on information gleaned from the huge volumes of data available on the Internet. The project expects to overcome one of the ....Deep visual understanding: learning to see in an unruly world. Deep Learning has achieved incredible success at an astonishing variety of Computer Vision tasks recently. This project aims to convey this success into the challenging domain of high-level image-based reasoning. It will extend deep learning to achieve flexible semantic reasoning about the content of images based on information gleaned from the huge volumes of data available on the Internet. The project expects to overcome one of the primary limitations of deep learning and will greatly increase its practical application to a range of industrial, cultural or health settings.Read moreRead less
Learning to Reason in Reinforcement Learning. Deep Reinforcement Learning (RL) uses deep neural networks to represent and learn optimal decision-making policies for intelligent agents in complex environments. However, most RL approaches require millions of episodes to converge to good policies, making it difficult for RL to be applied in real-world scenarios taking significant resources. This project aims to equip RL with capabilities such as counterfactual reasoning and outcome anticipation to ....Learning to Reason in Reinforcement Learning. Deep Reinforcement Learning (RL) uses deep neural networks to represent and learn optimal decision-making policies for intelligent agents in complex environments. However, most RL approaches require millions of episodes to converge to good policies, making it difficult for RL to be applied in real-world scenarios taking significant resources. This project aims to equip RL with capabilities such as counterfactual reasoning and outcome anticipation to significantly reduce the number of interactions required, improve generalisation, and provide the agent with the capability to consider the cause-effects. These improvements would narrow the gap between AI and human capabilities and broaden the adoption of RL in real-world applications.Read moreRead less
Intelligent Virtual Human Companions. This research aims to develop intelligent virtual human companions that can seemingly integrate our immediate physical environment and understand their surroundings including people’s emotions, behaviours, actions and interactions. Such a technology will be enabled by leveraging recent advances in mixed/augmented reality technologies, and by developing innovative artificial intelligence and computer vision and graphics algorithms for dynamic real-world envir ....Intelligent Virtual Human Companions. This research aims to develop intelligent virtual human companions that can seemingly integrate our immediate physical environment and understand their surroundings including people’s emotions, behaviours, actions and interactions. Such a technology will be enabled by leveraging recent advances in mixed/augmented reality technologies, and by developing innovative artificial intelligence and computer vision and graphics algorithms for dynamic real-world environments. Unlike robots, the proposed technology will be low cost, readily deployable and customisable, and will not have any physical limitations or maintenance requirements. It will thus have a wide range of applications from elderly care, healthcare care to educational training.Read moreRead less
Generative Visual Pre-training on Unlabelled Big Data. This project aims to develop a generative visual pre-training of large-scale deep neural networks on unlabelled big data. Developing pre-trained visual models that are accurate, robust, and efficient for downstream tasks is a keystone of modern computer vision, but it poses challenges and knowledge gaps to existing unsupervised representation learning. Expected outcomes include new theories and algorithms for unsupervised visual pre-training ....Generative Visual Pre-training on Unlabelled Big Data. This project aims to develop a generative visual pre-training of large-scale deep neural networks on unlabelled big data. Developing pre-trained visual models that are accurate, robust, and efficient for downstream tasks is a keystone of modern computer vision, but it poses challenges and knowledge gaps to existing unsupervised representation learning. Expected outcomes include new theories and algorithms for unsupervised visual pre-training, which are anticipated to deepen our understanding of visual representation and make it easier to build and deploy computer vision applications and services. Examples of benefits include modernising machines in manufacturing and farming with visual intelligence. Read moreRead less
A Novel Automatic Neural Network Feature Extractor. This project aims to study feature extraction abilities of convolutional as well as traditional neural networks and develop a generic feature extractor which can be applied to wide variety of real-world image and non-image data. New concepts for automatic feature extraction, feature explanation, hybrid evolutionary algorithms and non-iterative ensemble learning will be introduced and evaluated. The expected outcomes are a generic feature extrac ....A Novel Automatic Neural Network Feature Extractor. This project aims to study feature extraction abilities of convolutional as well as traditional neural networks and develop a generic feature extractor which can be applied to wide variety of real-world image and non-image data. New concepts for automatic feature extraction, feature explanation, hybrid evolutionary algorithms and non-iterative ensemble learning will be introduced and evaluated. The expected outcomes are a generic feature extractor for automatically extracting features, an optimiser for finding optimal parameters and non-iterative ensemble learning technique for classification of features into classes. The impact of this project will be automatic feature extractors and classifiers for real-world applications.Read moreRead less
Unlocking Mass Mobile Video Analytics with Advanced Neural Memory Networks. This project will develop neural memory architectures and dense spatial-temporal bundle adjustment to predict movement, behaviour, and perform multi-sensor fusion across large asynchronous video feeds. This capability will allow us to better interrogate and analyse mass video information recorded from the vast number of smartphones, action cameras, and surveillance cameras which exist at public events of interest. Outcom ....Unlocking Mass Mobile Video Analytics with Advanced Neural Memory Networks. This project will develop neural memory architectures and dense spatial-temporal bundle adjustment to predict movement, behaviour, and perform multi-sensor fusion across large asynchronous video feeds. This capability will allow us to better interrogate and analyse mass video information recorded from the vast number of smartphones, action cameras, and surveillance cameras which exist at public events of interest. Outcomes include the ability to ingest multiple video feeds into a dense and dynamic 3D reconstruction for knowledge representation and discovery, and analysis of events and behaviour through new spatio-temporal analytic approaches. This will offer significant benefits for video forensic analysis, policing, and emergency response.Read moreRead less