Towards a general framework for the annotation of dance motion sequences A framework and toolkit for collecting movement descriptions as ground-truth datasets

In this paper, we present a conceptual framework and toolkit for movement annotation. We explain how the design of the annotation systems, based on the framework, if combined with specific strategies for the process of annotation, can enhance the collection of ground-truth datasets for training algorithms. Computational algorithms, such as machine learning, show promising results for massive and scalable automatic movement annotation. Nevertheless, the need for reliable ground-truth datasets annotated by human experts, to train the machine learning algorithms and for bridging the gap between machine measurable and human perceived expressive aspects remains an open issue. This need constitutes a challenging task, due to the complexity of human movement and diversity of possible descriptors, as well as the high subjectivity that accompanies movement characterisation by both experts and non-expert users. We contribute to addressing this problem, by proposing a conceptual framework for dance movement manual annotation which we evaluate through the development and deployment of the toolkit. Finally, we discuss how the different design choices affect the process and the reliability of collecting data sets regarding qualitative aspects of movement.