Department of Computer Science

Our Research

Hand Orientation and Pose Estimation for Egocentric Devices

Recent trends in technology have resulted in the introduction of a number of egocentric devices that provide a first person view of the world around a person. However, the full potential of these devices is limited due to lack of natural interaction. Given the dynamic nature and limited resources on such devices, there is a need for a hand orientation and pose estimation method that can generalize well from different variations including orientation, hand pose, shape, size and style. Our research aims at developing novel machine learning models that can learn the mapping of 2D monocular hand images onto 3D hand orientation and pose. Our research in this domain has led to a number of machine learning methods that are capable of building good understanding of a given data despite the presence of variations arising from different sources.

Cone Counting in Adaptive Optics Imaging of the Retina

The purpose of the research work is to develop an automated computer vision system for the enhancement of AO high-resolution retinal images and quantitative assessment of photoreceptor cells. Enhancing high-resolution images of the retina will facilitate better distinction of photoreceptor cells, thus assisting the eye care professionals in the examination of the living retina, the diagnosis of different types of eye diseases and their early prediction. The proposed image processing framework consists of several stages: image quality assessment, illumination compensation, noise suppression, image registration, image restoration, image enhancement and cone detection. Within the framework, we tackle a specific noise component in the AO system, affecting the quality of acquired retinal data. Thus, we attempt to fully recover AO retinal images, free from any induced noise signals. A comparative study of different methods and evaluation of their efficiency on AO retinal datasets is performed by assessing image quality. In order to verify the achieved results, cone packing density distribution was calculated at different retinal locations and correlated with histological data. From the performed experiments, it can be concluded that the proposed image processing framework can effectively improve photoreceptor cell image quality and thus can serve as a platform for further investigations of retinal tissues.

Robust Hand Pose Estimation from Stereoscopic and Egocentric View Point

The research entails of hand gesture recognition from stereo-optically acquired depth data from egocentric viewpoints, with emphasis on applications such as sign language recognition, virtual/natural interaction, and gaming inputs amongst others. The potential of the research is to develop wearable stereo camera which is more suited to egocentric scenario where the relatively better form factor, energy consumption, near distance coverage, and outdoor usage provided by using RGB camera (compared to RGBD cameras) is essential. This resaerch naturally rests in the field of computer vision and machine learning, working with technologies such as Random forest, Eigenvectors, MCMC, Conditional Random Field.

Detection and Localization of Soft Plaques in Coronary CTA

Coronary atherosclerotic plaques are generally categorized into two categories including Calcified & Non-Calcified Plaques. NCPs are treated clinically more threatening due to high chances of unexpected rupture that results in fatal casualities. On the other hand, detection of NCPs is very difficult due to their apparent behaviour in CTA images as their intensity falls inbetween blood & fatty muscles. The focus of this research is to detect and quantify the soft plaques in coronary tree. The proposed researd starts with the levelset based segmentation of coronary tree from 3D CTA volume, followed with detection and localization of soft plaques in different segments. Luminal analysis based plaque quantification in the last stage can help clinicians in assessing the nature and extent of threat.

Cervical Vertebrae Segmentation and Injury Detection System for X-Ray Images

The cervical spine is a vital part of the human body, and due to its flexibility, is particularly vulnerable to trauma. Post-traumatic delayed or incorrect diagnosis can result in neurological deficit, paralysis and even death. Early and accurate detection of cervical spine injuries is critical to plan appropriate care and to prevent any tragic consequences. Despite standardisation and advances in imaging, missed or delayed diagnosis of cervical spine injuries in X-Ray images is still a common problem in emergency departments. Our goal is to develop a computer aided detection system based on the state-of-the-art advances in computer vision and machine learning that will help the radiologists to detect cervical bone injuries with high accuracy.

Pattern Classification of Neurological disorder based upon fMRI

The human brain is considered a vital part of our body responsible for all bodily functions. Neurological disorders have emerged as greatest threat to human wellbeing. Underlying mechanism of the most of the disorders is still not well understood and proper and early diagnosis is not available for most of the disorders. The aim of the research is to develop new machine learning models that can predict and differentiate between control and disordered subjects based upon functional MRI (fMRI) data.

Mesh Representations for Level of Detail in Animated Scenes

During the last two decades multi-resolution techniques have received much attention in the filed of computer graphics. Much of the existing multi resolution techniques has focused on techniques that involve static polygonal meshes. However, dynamically animated models has received less attention, and even less so for real time rendering. Furthermore, due to the increasingly highly complicated high resolution and realistic 3D graphics demanded by various industries, we identify an opportunity to do further research on this area. Our research aims to design and implement a state-of-the-art framework that will generate and select on-the fly the appropriate level-of-detail for dynamically animated models in a scene.

An Ontology-Based Framework for Image Understanding based on Visual Information gathered via Deep Learning

The use of convolutional neural networks has been vastly present in the latest generation of computer vision tools for object classification, detection and different object attributes recognition. However, for complete image understanding much more is required than just high performance in specific tasks against current benchmarks. Once the functional subsystems solving individual problems are developed they require a common knowledge base to be integrated, enabling semantic inferencing. The aim of this research project is to create an ontology for a bounded domain, which would serve as the core for a reasoning system that uses visual information extracted with convolutional neural networks.

Achievements

    Atif Riaz won the Three Minute Thesis competition held by City, University of London.
    Greg Slabaugh won the Research Student Supervision Award.
    The Computer Vision Group was awarded a Titan X Pascal GPU through NVidia’s GPU Grant Program.

News

    Paper "Probabilistic Spatial Regression using a Deep Fully Convolutional Neural Network" accepted to BMVC 2017.
    Anfisa’s and Asad's PhD viva was successful.
    Anfisa Lazareva is now an employee of Mirada Medical Ltd.