Transformer Performance Evaluation in 3D Reconstruction of Balinese Mask Wood Carving

This research aims to address the challenges in the preservation of Balinese mask wood carvings, which involve high costs, human resources, time, and technology. By applying transformer technology to 3D digitization, this study focuses on intricate details and accurate texture preservation. A research gap exists due to the lack of studies that have used transformer models for the 3D reconstruction of complex cultural art objects such as wood carvings. The methodology includes collecting 2D images of Balinese wood mask carvings, data pre-processing, transformer model training, and performance evaluation using the F1 score and Chamfer Distance. The novelty of this research lies in its use of transformer models in the context of a 3D reconstruction of cultural art objects for the first time. The primary contributions of this research include providing a new dataset and developing a model that can produce accurate 3D reconstructions for cultural preservation and its applications in the creative industry. The evaluation shows that although the Gandi class showed the best performance with an F1 score of 0.172829, precision of 0.247400, recall of 0.132800, and a Chamfer Distance of 0.007651, improvements are still needed. The Delem class demonstrated the lowest performance, which indicates the need for more advanced models. These results confirm the potential of transformer technology for digital preservation and offer a basis for the development of more efficient 3D reconstruction techniques to digitally document and preserve Bali's cultural heritage.