A5054998594- Ionosphere and magnetosphere dynamics
- Robotics and Sensor-Based Localization
- Solar and Space Plasma Dynamics
- Advanced Vision and Imaging
- Geomagnetism and Paleomagnetism Studies
- Geophysics and Gravity Measurements
- 3D Shape Modeling and Analysis
- 3D Surveying and Cultural Heritage
- Advanced Image and Video Retrieval Techniques
- Astro and Planetary Science
- Atmospheric Ozone and Climate
- Generative Adversarial Networks and Image Synthesis
- Human Pose and Action Recognition
- Robotic Path Planning Algorithms
- Earthquake Detection and Analysis
- GNSS positioning and interference
- Multimodal Machine Learning Applications
- Indoor and Outdoor Localization Technologies
- Computer Graphics and Visualization Techniques
- Industrial Vision Systems and Defect Detection
- Optical measurement and interference techniques
- Topic Modeling
- Anomaly Detection Techniques and Applications
- Underwater Vehicles and Communication Systems
- Radio Astronomy Observations and Technology
University of Oxford
2015-2025
Imperial College London
2017-2022
University of Virginia
2022
Xiaomi (China)
2022
University of Edinburgh
2022
University of the Witwatersrand
2013-2020
Dyson (United Kingdom)
2018-2019
National Solar Observatory
2008
Acuity Technologies (United States)
2006
Hewlett-Packard (United States)
2005
This paper studies monocular visual odometry (VO) problem. Most of existing VO algorithms are developed under a standard pipeline including feature extraction, matching, motion estimation, local optimisation, etc. Although some them have demonstrated superior performance, they usually need to be carefully designed and specifically fine-tuned work well in different environments. Some prior knowledge is also required recover an absolute scale for VO. presents novel end-to-end framework by...
The representation of geometry in real-time 3D perception systems continues to be a critical research issue. Dense maps capture complete surface shape and can augmented with semantic labels, but their high dimensionality makes them computationally costly store process, unsuitable for rigorous probabilistic inference. Sparse feature-based representations avoid these problems, only partial scene information are mainly useful localisation only. We present new compact dense which is conditioned...
In this paper we present an on-manifold sequence-to-sequence learning approach to motion estimation using visual and inertial sensors. It is the best of our knowledge first end-to-end trainable method for visual-inertial odometry which performs fusion data at intermediate feature-representation level. Our has numerous advantages over traditional approaches. Specifically, it eliminates need tedious manual synchronization camera IMU as well eliminating calibration between camera. A further...
We propose an online object-level SLAM system which builds a persistent and accurate 3D graph map of arbitrary reconstructed objects. As RGB-D camera browses cluttered indoor scene, Mask-RCNN instance segmentations are used to initialise compact per-object Truncated Signed Distance Function (TSDF) reconstructions with object size-dependent resolutions novel foreground mask. Reconstructed objects stored in optimisable 6DoF pose is our only representation. Objects incrementally refined via...
Machine learning techniques, namely convolutional neural networks (CNN) and regression forests, have recently shown great promise in performing 6-DoF localization of monocular images. However, most cases image-sequences, rather only single images, are readily available. To this extent, none the proposed learning-based approaches exploit valuable constraint temporal smoothness, often leading to situations where per-frame error is larger than camera motion. In paper we propose a recurrent...
This paper studies visual odometry (VO) from the perspective of deep learning. After tremendous efforts in robotics and computer vision communities over past few decades, state-of-the-art VO algorithms have demonstrated incredible performance. However, since problem is typically formulated as a pure geometric problem, one key features still missing current systems capability to automatically gain knowledge improve performance through In this paper, we investigate whether neural networks can...
In this paper, we propose a novel 3D-RecGAN approach, which reconstructs the complete 3D structure of given object from single arbitrary depth view using generative adversarial networks. Unlike existing work typically requires multiple views same or class labels to recover full geometry, proposed only takes voxel grid representation as input, and is able generate occupancy by filling in occluded/missing regions. The key idea combine capabilities autoencoders conditional Generative...
We propose a novel, conceptually simple and general framework for instance segmentation on 3D point clouds. Our method, called 3D-BoNet, follows the design philosophy of per-point multilayer perceptrons (MLPs). The directly regresses bounding boxes all instances in cloud, while simultaneously predicting point-level mask each instance. It consists backbone network followed by two parallel branches 1) box regression 2) prediction. 3D-BoNet is single-stage, anchor-free end-to-end trainable....
In this work, we propose a VAE-LSTM hybrid model as an unsupervised approach for anomaly detection in time series. Our utilizes both VAE module forming robust local features over short windows and LSTM estimating the long term correlation series on top of inferred from module. As result, our algorithm is capable identifying anomalies that span multiple scales. We demonstrate effectiveness five real world problems find method outperforms three other commonly used methods.
Datasets have gained an enormous amount of popularity in the computer vision community, from training and evaluation Deep Learning-based methods to benchmarking Simultaneous Localization Mapping (SLAM). Without a doubt, synthetic imagery bears vast potential due scalability terms amounts data obtainable without tedious manual ground truth annotations or measurements. Here, we present dataset with aim providing higher degree photo-realism, larger scale, more variability as well serving wider...
In this paper we present an on-manifold sequence-to-sequence learning approach to motion estimation using visual and inertial sensors. It is the best of our knowledge first end-to-end trainable method for visual-inertial odometry which performs fusion data at intermediate feature-representation level. Our has numerous advantages over traditional approaches. Specifically, it eliminates need tedious manual synchronization camera IMU as well eliminating calibration between camera. A further...
The ability to estimate rich geometry and camera motion from monocular imagery is fundamental future interactive robotics augmented reality applications. Different approaches have been proposed that vary in scene representation (sparse landmarks, dense maps), the consistency metric used for optimising multi-view problem, use of learned priors. We present a SLAM system unifies these methods probabilistic framework while still maintaining real-time performance. This achieved through compact...
Volume rendering using neural fields has shown great promise in capturing and synthesizing novel views of 3D scenes. However, this type approach requires querying the volume network at multiple points along each viewing ray order to render an image, resulting very slow times. In paper, we present a method that overcomes limitation by learning direct mapping from camera rays locations are most likely influence pixel's final appearance. Using able render, train fine-tune...
This paper studies indoor localisation problem by using low-cost and pervasive sensors. Most of existing algorithms rely on camera, laser scanner, floor plan or other pre-installed infrastructure to achieve sub-meter sub-centimetre accuracy. However, in some circumstances these required devices information may be unavailable too expensive terms cost deployment. presents a novel keyframe based Pose Graph Simultaneous Localisation Mapping (SLAM) method, which correlates ambient geomagnetic...
In this paper, we present a novel end-to-end learning-based LiDAR sensor relocalization framework, termed PointLoc, which infers 6-DoF poses directly using only single point cloud as input. Compared to visual sensor-based relocalization, sensors can provide rich and robust geometric information about scene. However, clouds of are unordered unstructured making it difficult apply traditional deep learning regression models for task. We address issue by proposing PointNet-style architecture...
Abstract. The basic aim of this ‘case study’ is to investigate the variability in maximum height ionospheric F2-layer, hmF2, with periods planetary waves (2–30 days), and make an attempt determine their origin. hourly data hmF2 above Millstone Hill (42.6° N, 71.5° W) during 01 September 1998 - 31 March 2000 were used for analysis. Three types disturbances are studied detail: (i) 27- day oscillations observed generated by geomagnetic activity global-scale 27-day wave present zonal...
Scaling test-time compute is a promising axis for improving LLM capabilities. However, can be scaled in variety of ways, and effectively combining different approaches remains an active area research. Here, we explore this problem the context solving real-world GitHub issues from SWE-bench dataset. Our system, named CodeMonkeys, allows models to iteratively edit codebase by jointly generating running testing script alongside their draft edit. We sample many these multi-turn trajectories...