Historical Baroque dance, a cornerstone of European court culture primarily spanning the late 17th and early 18th centuries, represents a unique challenge for cultural heritage preservation. Unlike static arts, dance is ephemeral. While the era left behind rich resources—specifically handwritten manuals and the complex Beauchamp-Feuillet notation—interpreting these sources requires highly specialized knowledge. The difficulty of deciphering historical handwriting combined with abstract choreographic symbols creates a significant accessibility threshold, isolating this art form from the public and endangering its transmission as Intangible Cultural Heritage (ICH).

This project proposes a comprehensive Digital Humanities pipeline utilizing AI to bridge the gap between archival sources and modern visual understanding. The framework begins with Optical Character Recognition (OCR) tailored for historical handwriting, allowing for the digitization and interpretation of original dance manuals. This textual and symbolic data then forward together with videos into a large model designed to detect and generate Baroque dance sequences in 3D, rendering the movements accessible for educational and archival purposes.

Technically, we employ a two-stage approach for 3D human pose estimation. While end-to-end methods offer efficiency, our research prioritizes the higher accuracy inherent in two-stage approaches, which first detect 2D keypoints and subsequently lift them to 3D coordinates. To mitigate the scarcity of historical dancers and motion capture resources, we focus initially on solo dance sequences. We address the risk of overfitting through rigorous data augmentation and transfer learning techniques.

A critical contribution of this research is the creation of a domain-specific dataset. Current state-of-the-art datasets for human pose estimation—such as Human3.6M, MPI-INF-3DHP, and EMDB—focus primarily on daily human actions. While specific dance databases like AIST and AIST++ exist, they lack the specific stylistic nuances, distinct postures, and ornamental gestures unique to the Baroque period. To fill this lacuna, we are collaborating with professional Baroque dancers to capture and annotate a dedicated dataset, establishing a ground truth that links historical notation to biomechanical reality.

By converting static, handwritten historical records into dynamic 3D visualizations, this pipeline offers an all-encompassing view of Baroque performance. This not only lowers the barrier for public appreciation and amateur learning but also provides a scalable digital framework applicable to other forms of historical dance. Ultimately, this project demonstrates how computational methods—from OCR to 3D reconstruction—can serve as robust tools for protecting, preserving, and democratizing access to our shared, intangible cultural history.