Home > Research Areas

Research Areas

AI's main goal is to develop adaptive tools for understanding and using complex natural signals, particularly images of everyday scenes. Our recent work has centred mainly on the following three areas:

  1. Visual recognition and scene understanding. We have particular expertise in methods for detecting everyday objects (people, cars, animals, etc) in images of everyday scenes, and in segmenting such images into semantically meaningful regions (regions of buildings, foliage, water, etc). We also work on human face recognition.
  2. Feature extraction for computer vision and signal processing. We have a good deal of experience in the development of robust features for visual recognition. Our "Histogam of Oriented Gradients" features for object detection (developed before AI separated from LEAR in 2005-2006) have been used and cited many hundreds of times during the current period, and we continue to work on creating even more robust and discriminant feature sets.
  3. Machine learning and statistical modelling of signals. We use a wide variety of machine learning and statistical tools, including various kinds of classical and machine learning discriminants, weakly supervised and latent variable methods. Our focus is on developing methods and algorithms that are able to handle the kinds of noisy, high-dimensional problems with weak labelling and massive data sets that commonly arise in computer vision and natural signal processing, but there is also a small amount of more basic work, notably on discrimination using convex models, and on the relationships between discriminative and generative methods and between machine learning and statistical formulations.

In the future, these themes will continue but the range of signals that is treated is likely to be widened. We have already begun to work with remote sensing imagery (currently SAR) for environmental applications, and depending on future recruitments work on areas such as audio or medical imaging is also a possibility.