A Stereoscopic-Feature Prior for Enhancing 3D Digitization of Pose Variant Terracotta Figurines

Deep learning-based 3D reconstruction technology is an active research area in the field of cultural relic restoration. To address the issue that existing reconstruction algorithms cannot accurately segment moving parts with significant deformation and self-occlusion, we propose an implicit function based on prior stereo features, which achieves explicit modeling of pose variations through stereo feature extraction and pixel feature extraction. Firstly, the stereo feature extraction module DenseUNet++ is obtained by combining an improved U-Net and a dense connected network. A Texture front normal image is added to the input end to provide stereo priors for the network. Subsequently, pixel feature extraction uses a parallel dual-branch network to extract image pixel features and combines stereo and pixel features to construct an implicit function for single-view reconstruction. Additionally, we construct 3D terracotta datasets (VPTerracotta Datasets) that contain rich motion and high-precision details. Finally, we conducted a large number of comparative demonstrations on the self-built VPTerracotta datasets and the public RenderPeople dataset. The experimental results show that our method can effectively reconstruct the occluded parts when dealing with self-occluding objects, and the proposed reconstruction model exhibits good coherence and a certain degree of universality.