Digital Twin-Based Stereo Dataset Generation Method for 3D Reconstruction of Teeth

Producing stereo matching datasets in dental 3D reconstruction studies is expensive and difficult. In this study, we propose a method for procedurally generating high-precision virtual stereo matching datasets based on the Unity engine. By accurately modeling the binocular camera imaging principle and environmental parameters, the method can generate paired stereo images with known depth information (disparity maps), sampling depth values with an accuracy of up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10^{-4} ~\mathrm{m}$</tex>. This method breaks through the traditional paradigm of relying on expensive sensor acquisition or limited public datasets. It effectively solves the traditional problems of difficulty in acquiring depth information of real scenes, scarcity of high-quality datasets and high cost. The experiments were performed on the widely used open-source stereo matching model RAFT-Stereo and aligned to the target tooth point cloud after segmentation. The results show that the performance of the model trained based on this dataset is comparable to the traditional real dataset. The average alignment error for a single tooth was <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{0. 1 7 - 0. 2 6 ~ m m}$</tex>, with the maximum error below 0.93 mm. For multiple tooth scenarios, the average error of multiple co-alignment was stabilized at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{0. 2 7 - 0. 7 9 ~ m m}$</tex>. Finally, experiments were performed on facial images taken by a real camera. The average error of the reconstructed point cloud is within 1 mm. The potential of virtual data as a reliable alternative to real data is verified, and the scheme provides a convenient, low-cost and controllable data source for the study of stereo matching algorithms.