Texture-Semantic Point: Registration for Point Clouds of Porcelain Relics

Point cloud models of porcelain are captured through the multi-registration of point clouds, which presents a challenging task. On the one hand, the smooth surface of porcelain lacks geometric feature variation, making it difficult to establish corresponding points. On the other hand, the overall geometric symmetry of most porcelain relics can easily lead to iterative calculations falling into the local minimum convergence trap. To address the difficulties in feature point selection, we propose a novel approach using texture-semantic points as features for coarse registration. We first select rich texture points as 2D candidates and establish a 2D-3D matching relationship, giving each candidate its 3D spatial location and associated texture information. Using these correspondences, we perform coarse alignment of the point clouds. However, in reality, the point clouds are not aligned, and the registration calculation fails because of geometric symmetry issues. To address this, we integrate a control net into the iterative closest point (ICP) calculation to guide iterations towards the correct Special Euclidean group in 3 dimensions (SE(3)) transformation, achieving refined alignment. Finally, considering porcelain's symmetrical geometry, we introduce a pose optimization constraint using the symmetry axis as a weighted parameter to limit degrees of freedom and enhance registration accuracy. Experiments were conducted on seven porcelain datasets to evaluate the proposed approach. A qualitative analysis demonstrated successful refined alignment using the proposed approach. In addition, we performed a quantitative comparison with state-of-the-art methods. Experimental results showed that our approach outperformed others across all models when applied to the registration of geometrically symmetric porcelain; Specifically, the proposed method achieved a 50% enhancement in accuracy compared with others, measured by the distance between the labeled corresponding points.