Occlusion-aware and consistency-enhanced 3D Gaussian splatting for crowdsourced heritage reconstruction

In recent years, 3D Gaussian Splatting (3DGS) has shown great potential for high-fidelity rendering in the digital preservation of cultural heritage. However, for immovable cultural relics, the image acquisition process is often uncontrollable, resulting in numerous occlusions that leads to artifacts and texture distortions. This paper presents an occlusion-aware 3DGS optimization framework incorporating multi-view photometric consistency. Occlusion masks are generated via object detection and used to guide image inpainting, producing fused occlusion-aware inputs for training. A region-wise loss function is designed to preserve visible textures, while the inpainted content guides the occluded regions toward perceptual realism. Moreover, a multi-view photometric consistency loss enforces appearance coherence across views by projecting reference images into target viewpoints. Experimental results demonstrate that the proposed method effectively reduces occlusion-induced artifacts and significantly improves reconstruction accuracy and visual consistency on crowdsourced heritage datasets, highlighting its strong practical value and potential for broader applications.