Activity Recognition from Sparsely Labeled Data Using Multi-Instance Learning

Maja Stikic, Bernt Schiele

Abstract: Activity recognition has attracted increasing attention in recent years due to its potential to enable a number of compelling context-aware applications. As most approaches rely on supervised learning methods, obtaining substantial amounts of labeled data is often an important bottle-neck for these approaches. In this paper, we present and explore a novel method for activity recognition from sparsely labeled data. The method is based on multi-instance learning allowing to significantly reduce the required level of supervision. In particular we propose several novel extensions of multi-instance learning to support different annotation strategies. The validity of the approach is demonstrated on two public datasets for three different labeling scenarios.