Preserving and Annotating Dance Heritage Material through Deep Learning Tools: A Case Study on Rudolf Nureyev

The cultural heritage of theatrical dance involves diverse sources requiring complex multi-modal approaches. Since manual analysis methods are labor-intensive and so limited to few data samples, we here discuss the use of the DanXe framework, which combines different AI paradigms for comprehensive dance material analysis and visualization. However, DanXe lacks models and datasets specific to dance domains. To address this, we propose a human-in-the-loop (HITL) extension to the DanXe to accelerate multi-modal data labeling through semi-automatic, high-quality data labeling. This approach aims to create detailed datasets providing humans with a set of user-friendly and effective tools for advancing multi-modal dance analysis and optimizing AI methodologies for dance heritage documentation. To this date, we designed a novel middleware that allows us to adapt data generated from visual Deep Learning (DL) models within DanXe to visual annotation tools, to empower domain experts with a user-friendly tool to preserve all the components included in the choreographic creation, enriching the process of metadata creation.