Computer vision models predict where objects are in an image, and what those objects are. These vision models give robots the ability to perceive their environment and choose safe and smart actions based on this perception.
Computer vision models can fail silently when exposed to unexpected or difficult environments - e.g. changes in camera viewpoints, changes in lighting, or when seeing new objects that haven't been seen before. This raises concerns about the safety of using vision models in the real world or on board robots.
In this project, you'll explore where and when computer vision models fail. You'll have the opportunity to learn about deep learning and computer vision and the current research frontiers in these popular topics. This will allow you to gain invaluable hands-on experience with state-of-the-art computer vision models and deep learning tools.
If you're interested in experiencing the world of research and upskilling yourself in computer vision and deep learning, this project is for you.
Your research activities will include:
- survey the literature on failure and failure detection in computer vision
- collect data and experiment with state-of-the-art computer vision models
- develop techniques to detect when computer vision fails.
If interested, you can also be involved in:
- writing academic research papers for submission at top tier academic conferences and journals
- collaboration with fellow students, postdocs and academics.
The project aims to:
- develop a categorisation of failure modalities in computer vision models
- produce visualisations and documentation of where computer vision models fail
- review existing techniques that address failures in computer vision
- develop new techniques for addressing failures in computer vision.
Skills and experience
You must have experience with coding/programming (particularly in python).
Ideally you will also have experience and skills (or the interest to gain skills) in:
- machine learning
- deep learning
- computer vision
- image processing.
Contact the supervisor for more information.