Active Class Selection for Dataset Acquisition in Sign Language Recognition

Dataset collection for Sign Language Recognition (SLR) represents a challenging and crucial step in the development of modern automatic SLR systems. Typical acquisition protocols do not follow specific strategies, simply trying to gather equally represented classes. In this paper we provide some empirical evidences that alternative, more clever, strategies can be really beneficial, leading to a better performance of classification systems. In particular, we investigate the exploitation of ideas and tools of Active Class Selection (ACS), a peculiar Active Learning (AL) context specifically devoted to scenarios in which new data is labelled at the same time it is generated. In particular, differently from standard AL where a strategy asks for a specific label from an available set of unlabelled data, ACS strategies define from which class it is more convenient to acquire a new sample. In this paper, we show the beneficial effect of these methods in the SLR scenario, where these concepts have never been investigated. We studied both standard and novel ACS approaches, with experiments based on a challenging dataset recently collected for an ECCV challenge. We also preliminary investigate other possible exploitations of ACS ideas, for example to select which would be, for the classification system, the most beneficial signer.