Thangka images present a challenging single-view 3D reconstruction problem: scattered-point perspective, dense overlap, and fine ornamentation often give rise to topological adhesion, contour distortion, and loss of high-frequency details. We propose GeoThangka-Net, a semantic-guided, topology-aware framework for heritage-oriented 3D digitization of Thangka principal deities. The method combines semantic scene decoupling and multimodal prior generation with an adaptive occlusion-aware visibility field for manifold truncation near occlusion boundaries, and a refinement across multiple scales with frequency-aware supervision, coupling monocular depth anchoring with multi-scale frequency-aware refinement. On a cross-domain geometric proxy dataset with 420 topologically complex samples, GeoThangka-Net improves PSNR by 2.32 dB and reconstruction stability by 25.6% over representative baselines, while also improving complementary 3D geometric agreement. Human perceptual evaluation further indicates better structural plausibility. The resulting reconstructions serve as heritage-oriented geometric surrogates for visualization, interaction, and comparative analysis.
Guo et al. (2026) studied this question.