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

The evolution of machine learning techniques has enabled significant advances in the field of object detection using computer vision applied to historical buildings. Indeed, studies in this area can contribute to applications ranging from automatic digital documentation to the technological enhancement of inspections in cultural heritage sites. However, a recurring challenge lies in the quality and public availability of image datasets that enable the application of deep learning models in the context of heritage churches, particularly those focused on the interior environments of these religious buildings. In this context, the objective of this paper is to present a novel annotated dataset for object detection in Brazilian heritage altars using deep learning. The proposed dataset, named AltarData, comprises 1278 high-quality images collected from 33 churches located in 12 historical cities in Brazil. Three of these urban ensembles are recognized as UNESCO World Heritage Sites: the Historic Town of Ouro Preto, the Historic Centre of Salvador de Bahia, and the Historic Centre of the Town of Diamantina. The data were carefully annotated, resulting in 12,093 labeled instances across six categories frequently found in Baroque and Rococo heritage altars: altar table, column, crucifix, saint sculpture, tabernacle, and throne. In addition, a deep learning with hyperparameter tuning approach was proposed, enabling the optimization of object recognition in heritage altars, achieving validation accuracies of up to 92.0% (altar table), 87.1% (column), 72.2% (crucifix), 85.1% (saint sculpture), 97.3% (tabernacle), and 84.7% (throne). In the test experiments, an F1-score of 79% was achieved using the MobileNetV2 configuration with width multiplier 0.35 and learning rate 0.0100. Overall, this paper contributes to the research field of new datasets for deep learning in cultural heritage sites, fostering advances in multi-object recognition within the interior environments of historical churches.