3D Single-Tooth Reconstruction via Curriculum Learning and Topology-aware Generation.

Abstract

Automated single-tooth reconstruction is a pivotal task in modern digital dentistry, yet existing point cloud completion methods often lack the specialized anatomical priors and spatial constraints necessary for clinical integration. In this paper, we propose Tooth-STR, a hierarchical generative framework designed to synthesize anatomically high-fidelity and collision-free tooth models from partial intraoral scans. Our Tooth-STR consists of a Topology-Aware Coarse Generation (TACG) module that utilizes differential edge convolutions to reason about the missing dental manifold based on the topological trends of the surrounding dentition, and a Curriculum Anatomical Constraint (C$^{2}$) strategy that implements a phased, differentiable boundary penalty to balance unconstrained geometric exploration with strict spatial regularization. Extensive experiments on the constructed Tooth-STR dataset demonstrate that our method significantly improves reconstruction performance across various tooth categories.