Simulating user intervention for interactive semantic place recognition with mobile devices

Recognising and learning users' semantically meaningful places is useful for personalising services and recommender systems, particularly in a mobile environment. Existing approaches that use mobile devices focus on automating place inference from underlying data, i.e. with little user interaction and intervention -- where user feedback is incorporated into the inference process. The process of intervention can be burdensome to the user but, without intervention, it is difficult to both capture personal place semantics and update places over time; resulting in a trade-off between system performance and user burden. In this paper, we present early results from a place recognition and learning approach that relies on user intervention as a form of active learning. Using simulations of user intervention generated from fine-grained ground truth, we show that good place semantic capture, classification and learning performance can feasibly be achieved in real time on mobile devices with only a small amount of user intervention.