The Interaction Behavior Dataset: A Dataset of Smiles and Laughs in Dyadic Interaction

We present a dataset of smiles and laughs in dyadic interactions as well as other information, namely their respective intensities and the participants dialog turns (roles). The motivation behind building this dataset is to contribute to the development of interactive virtual agents for which nonverbal expressions, although being very important components of communication, are neglected mainly due to the lack of resources. Available datasets tend to neglect at least one of the annotation data that we provide here. We first present an annotation protocol designed to optimize the annotation time and minimize the subjectivity in the annotation. Indeed, given the subjective nature of the annotation process of such expressions, reducing the subjectivity would allow to leverage several annotators while keeping the annotations homogeneous. We present measurements of the annotation protocol's efficiency. The dataset is available here https://github.com/numediart/ibdataset.