Continously updating place descriptors

Visual place recognition refers to the ability to recognise previously visited places even under changing conditions. Most place recognition methods aim to extract universal descriptors that enable to re-localise under drastic appearance and viewpoint changes. In this project, you will instead explore whether we can borrow concepts from the continual learning literature to update the place descriptors over time. The aim is to detect the domain shift and adapt the descriptor depending on the nature of the change – for example, the onset of rain requires fast adaptation, while a change in season requires slow adaptation. Simply adding more and more reference images is detrimental, as compute and storage requirements increase, and potentially more visually aliased images are added. Recent works on continual learning with spiking networks have shown exciting advances that this project will extend and adapt to the place recognition scenario.

Chief Investigators