Module-wise Adaptive Distillation for Multimodal Foundation Models
Multi-armed bandits for adaptive layerwise distillation of large pre-trained foundation models.
I am a Research Scientist at Google.
Previously, I was an Associate Professor (Reader) at the University of Bath (2011-2015), Founder and CTO of Cloudburst Research Inc. (2009-2015), and a Postdoctoral Researcher at EPFL (2009-2011), UBC (2008-2009) and Microsoft Research (2006-2007).
My interests are in Computer Vision, Machine Learning and Environmental Informatics. In 2017 I won the ICCV Helmholtz Prize (test of time award) for my work with David Lowe on panoramic image stitching.
Multi-armed bandits for adaptive layerwise distillation of large pre-trained foundation models.
Given a lot of images of an object category, you can train a NeRF to render them from novel views and interpolate between different instances.
MoViNets are a family of efficient video classification models supporting frame-by-frame inference on streaming video.
Augment classic class-balanced learning by estimating differences between class-conditioned distributions via meta-learning
Create datasets and study the effect of non-identical data distribution on Federated visual classification.
Initial work on non-IID data distribution for visual classification via Federated Learning
Recognise novel categories by imprinting final layer weights using the activations from one or more exemplars.
End-to-end trainable frame-recurrent video super-resolution framework that uses the previously inferred HR estimate to super-resolve the subsequent frame
Learn segmentation without ground truth by playing a game of cut and paste in an adversarial learning setup
SFMLearner: Joint learning of monocular depth plus camera pose from image streams without camera or depth supervision.
Building energy modelling and event detection using kernel regression
New RGB-NIR scene dataset and multispectral SIFT variant
Learn local descriptors by sampling corresponding image patches from large scale 3D reconstructions.
Location recognition in large image sets using informative feature vocabulary trees
Fully automatic structure and motion for unordered image sets.
Improved version of the AutoStitch panorama stitcher
First solution for panoramic stitching without user input
This course provides an introduction to the fundamental principles and applications of computer vision.
A masters course in computer vision.