This study proposes a four-stage pipeline for the individual 3D reconstruction of overlapping artifacts depicted in Korean Royal Court Paintings. Korean Royal Court Paintings are a valuable visual record of historical material culture, but the occlusi...
This study proposes a four-stage pipeline for the individual 3D reconstruction of overlapping artifacts depicted in Korean Royal Court Paintings. Korean Royal Court Paintings are a valuable visual record of historical material culture, but the occlusion caused by their overlapping compositions hinders the complete assetization of individual objects. To address this, the proposed pipeline consists of (1) Image Restoration, (2) Object Segmentation, (3) Inpainting (SSEI), and (4) 3D Reconstruction, with each stage connected through explicit input-output interfaces so that it can be replaced independently. The Image Restoration stage sequentially applies three sub-steps: Split-Radius Spectral Interpolation, which removes the weave pattern in the frequency domain; Spatial-Adaptive NLM, which locally smooths the residual weave pattern and chromatic components remaining near object boundaries and in narrow regions; and Contour Enhancement, which restores the weakened contour contrast. The Object Segmentation stage refines coarse LabelMe polygon annotations through SAM-based two-pass constrained refinement (shrink-only and interior preservation) and extracts each object as an RGBA image. The Inpainting (SSEI) stage first detects occluded regions via label-based layer ordering, and then restores them using a training-free self-exemplar approach (Style-consistent Self-Exemplar Inpainting, SSEI) that draws reference patches only from the visible region. Here, patches near the occlusion boundary of the same object are filled progressively from the boundary inward, a multi-stage color-consistency check suppresses the intrusion of incongruent colors, and the shared cross-boundary line is restored at a variable width matched to the stroke thickness measured in the visible region, thereby reviving the occluded boundary line of the object. The final 3D Reconstruction stage converts each completed individual-object image into an independent 3D textured mesh.
SSEI, the core contribution, is a training-free approach in which every synthesized pixel is copied from the visible region of the original painting itself, so that the algorithm structurally guarantees style preservation. In an evaluation on 50 synthetic occlusion cases extracted from Korean Royal Court Paintings, SSEI ranked first across all three metrics, leading with a PSNR of 19.25 dB alongside the best SSIM and LPIPS scores. While the two classical interpolation algorithms, Navier-Stokes and Telea, formed the next tier at around 15 dB, all learning-based models—generative methods PowerPaint and BrushNet as well as LaMa and SD-Inpaint—fell short of SSEI; in particular, several of these models trained on natural images remained at the 9-10 dB level, even below the classical algorithms, revealing the limited cultural-heritage-domain transferability of learning-based models and suggesting that domain distribution mismatch is a more dominant determinant of performance than architectural sophistication. This advantage reflects a robust performance gap rather than a mere difference in means: it was statistically significant over all baselines in the Wilcoxon signed-rank test (p 〈 0.01), SSEI was clearly separated from the next-ranked method with the lowest mean rank in the critical difference diagram, and every comparison fell within the large-effect region in the Cohen's d effect-size analysis.
In conclusion, this study establishes a self-exemplar-based inpainting paradigm that quantitatively preserves stylistic consistency in the cultural-heritage domain where training data are scarce, systematizes a heritage-specific preprocessing methodology, and integrates them into an end-to-end pipeline, thereby presenting a systematic approach for converting individual artifacts in the flatness-emphasizing, scattered-perspective Korean Royal Court Paintings into independently manipulable 3D digital assets. The proposed pipeline holds potential for applications such as the 3D assetization of elements in court paintings, virtual content production, and hypothetical visualization of occluded artifacts. In doing so, this study provides the first empirical evidence for the validity of a style-consistent, self-exemplar-based approach that overcomes the limitations of learning-based state-of-the-art models for the unsolved problem of 3D reconstruction of occluded objects in Korean Royal Court Paintings, located at the intersection of cultural heritage and computer vision, and it is hoped that the proposed SSEI and integrated pipeline will serve as a reference point for future research on the digital restoration of Korean painting heritage.