AltarData: A novel annotated dataset for object detection in the Brazilian heritage altars using deep learning

A novel annotated dataset, AltarData, achieves high accuracy in object detection within Brazilian heritage altars using deep learning, with validation accuracies reaching up to 97.3% for tabernacles and an F1-score of 79% in test experiments. This breakthrough addresses a critical gap in available datasets for cultural heritage research, enabling advancements in automatic digital documentation and technological inspections of historical churches. The dataset comprises 1278 high-quality images from 33 churches across Brazil, including three UNESCO World Heritage Sites, with annotations covering six key altar components: altar table, column, crucifix, saint sculpture, tabernacle, and throne. A deep learning approach with hyperparameter tuning was employed to optimize object recognition, demonstrating the potential for multi-object detection in interior environments of historical religious buildings.

Ottoni, A. L. C., & Ottoni, L. T. C. (2026). AltarData: A novel annotated dataset for object detection in the Brazilian heritage altars using deep learning. Journal of Cultural Heritage, 81, 248–259. https://doi.org/10.1016/j.culher.2026.07.024