Transformer-Based Single Image 3D Reconstruction for Cultural Artifacts: A Case Study on Balinese Masks

The digital preservation of cultural artifacts is essential for heritage archiving and public engagement. This paper introduces a novel single image 3D reconstruction framework specifically for these objects, demonstrated on the complex topology of Balinese masks. Our approach unifies a Transformer based global shape prior to capture holistic structure from 2D inputs with SDF diffusion refinement to recover high frequency details and watertight topology. On BaliMask3D, we establish a reproducible single image protocol spanning synthetic rendering, strict data splits, and multi aspect metrics: Chamfer Distance (CD), Uniform Hausdorff Distance (UHD), Intersection over Union (IoU), Topological Morphology Distance (TMD), and F1 Score on a downstream classification task. Key results reveal consistent gains and a clear trade off pattern: exemplar cases reach F1 Score 0.9676 with CD 0.0159; UHD as low as 0.0259 (e.g., Lenda) and TMD 0.018 indicate strong topological preservation, while volumetrically complex masks (e.g., Dewi Sita, Hanoman) highlight superior volumetric overlap (IoU) despite extremely fine surface details. Ablation studies validate the complementary roles of the global Transformer prior and local diffusion based refinement. The framework enables cost effective 3D digitization from a single photograph for archiving, virtual museums, and public engagement. We release protocols, evaluation scripts, and model artifacts to support reproducibility and include ethical and cultural sensitivity considerations for data creation and distribution.