Visual Artificial Intelligence Laboratory (VAIL)

 News

The paper

ROAD-R: The Autonomous Driving Dataset with Logical Requirements 

by Eleonora Giunchiglia, Mihaela Stoian, Salman Khan, Fabio Cuzzolin and Thomas Lukasiewicz has won the Best Paper Award at the IJCAI 2022 Workshop on Artificial Intelligence for Autonomous Driving (AI4AD 2022)

The extended version of the paper has been accepted by Machine Learning journal, and has won the Best student paper prize at IJCLR 2022, the International Joint Conference on Learning and Reasoning.

About us

The Visual Artificial Intelligence Laboratory was founded in 2012 by Professor Cuzzolin under the name of 'Machine Learning' (and later 'Artificial Intelligence and Vision') research group, and has since conducted work at the boundaries of human action recognition in computer vision. Prof Cuzzolin is a leading scientist in the mathematics of uncertainty, in particular random set and belief function theory.

Our research interests span a number of frontier topics in:

  • computer vision (action and activity detection, future event prediction, video captioning and scene understanding)
  • machine learning (continual learning, federated learning, self-supervision and metric learning)
  • artificial intelligence (epistemic AI and machine theory of mind, but also neurosymbolic AI),
  • robotics (with a focus on surgical robotics), autonomous driving (the detection of road events for situation awareness)
  • AI for healthcare (the monitoring of people in care homes, the early diagnosis of dementia, empathetic healthcare via theory of mind).

More information about VAIL

Related courses

The laboratory has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreements No. 964505 (E-pi) and No. 779813 (SARAS).

Research impact

Road event detection in autonomous driving, with colored boxes around the relevant road agents to be detected

The group has built, in just a few years, a leadership position in the field of deep learning for action detection, with some of the best detection accuracies to date and the first ever system able to localise multiple actions on the image plane in (better than) real time. The team's effort is now shifting towards topics at the frontier of computer vision, such as future action prediction, deep video captioning and the development of a theory of mind for machines.

The Lab currently runs on a budget of around £3.2M (not fully incorporating the €4.3M Horizon 2020 project SARAS or the €3M FET Epistemic AI we are coordinating), with currently nine live projects funded by Horizon 2020, the Leverhulme Trust, Innovate UK, Huawei Technologies, UKIERI, and the School of Engineering, Computing and Mathematics. The budget is projected to further significantly increase in 2022.

Prof Cuzzolin's reputation in uncertainty theory and belief functions comes from the formulation of a geometric approach to uncertainty in which probabilities, possibilities, belief measures and random sets are represented and analysed by geometric means. This has recently developed into an effort to reshape the foundations of artificial intelligence to better incorporate and model second-order, 'epistemic' uncertainty: an approach that we call Epistemic Artificial Intelligence.

Leadership

Fabio Cuzzolin

Professor Fabio Cuzzolin

Professor of Artificial Intelligence

View profile

Membership

slide 1 of 3

Projects

slide 1 of 2

Resources

Below you can find links to a number of resources generated by our research, including datasets and code.

ROAD is the ROad event Awareness Dataset for autonomous driving, released at the ROAD @ ICCV 2021 workshop.

Our ICCV'17 code on real-time action detection, the first online solution ever published.

The Continual Activity Recognition (CAR) dataset was released at the CSSL @ IJCAI 2021 workshop.

3D RetinaNet is our event detection approach, used as baseline for detection tasks in the ROAD dataset.

The Continual Crowd Counting (CCC) dataset was released at the CSSL @ IJCAI 2021 workshop.

Avalanche: the End-to-End Library for Continual Learning created by our partners ContinualAI.

The SARAS-MESAD dataset is a surgical action detection dataset released at MICCAI 2021 as part of the SARAS project.

Code for the BMVC 2018 paper 'Incremental Tube Construction for Human Action Detection'.