What differs? - Deep Learning for Anomaly and Out of Distribution Detection in Computer Vision (Visual AI) Applications

About the Project

Anomaly detection, and the synonymous topics of novelty and out-of-distribution detection, addresses the computer vision ("visual AI") challenge of automatically identifying outliers within the scene or image based on their appearance, motion or behaviour characteristics. Essentially asking - "what differs and why?" - but doing this algorithmically in the age of Large Language Models (LLM) and Vision Language Models (VLM) tools.

This represents an important and application-relevant challenge within both computer vision and the broader field of artificial intelligence that can be applied both across a wide range of data sources (e.g. images/video, 3D LiDAR/radar/point cloud data, bio-signals etc) and the imaging spectrum (visible, thermal infrared, X-ray etc). This project aims to focus on the use of deep learning based computer vision approaches for this task, with a particular focus on developing approaches that can address the challenges of dataset imbalance, continuous learning and/or complex scene contexts.

There are a wide range of applications for automatic anomaly and out of distribution detection approaches, that could form the basis for a specific PhD project:

• on-line learning for continuous automated wide area surveillance in CCTV applications
• 2D and 3D security X-ray image screening for transportation and border security applications
• outlier detection within vehicle perception systems for future autonomous road vehicles
• automated wide area search and detect for automating future drone-based search and rescue operations
• automated in-situ infrastructure inspection using varying sensor capabilities mounted on autonomous robotics and/or drones