I am excited about research which advances the perception and control of mobile robotics. In particular, I am currently working on improving the robustness and accuracy of computer vision algorithms, leveraging geometry for self-supervised learning and developing end-to-end systems which can reason from perception to control.
Reasoning from Perception to Action
The computer vision systems which I develop are primarily motivated to extract a representation which can be used to make a decision or action. I’m interested in learning representations to understand scenes and control the behavior of real world robots. For example, we designed a reinforcement learning agent which can learn to drive a car with deep reinforcement learning.
Scene understanding is a fundamental task in computer vision which requires understanding information such as the scene’s semantics, geometry and motion. Initially, I worked on a semantic segmentation algorithm called SegNet. More recently, I have been interested in learning a representation from a multitask deep learning architecture.
Bayesian Deep Learning
Deep learning is great for achieving state-of-the-art results, however these models cannot understand what they don’t know. Bayesian deep learning (BDL) is a very exciting framework for understanding our model’s uncertainty. This paper is an introduction to Bayesian deep learning for computer vision. I have also found BDL useful for localisation, scene understanding and autonomous driving.
PoseNet is an algorithm for relocalisation - estimating the position and orientation of the camera from an image within a previously explored area. It works over large outdoor urban environments or inside buildings. It takes only 5ms to do this from a single colour image, here is a demo.
Some more details of other projects, including an autonomous drone and augmented reality, can be found here.