Bohdan Vodianyk¹, Enrique Nava Baro¹, Alfonso Ariza Quintana¹ and Anton Popov², ¹Universidad de Málaga, Spain, ²Igor Sikorsky Kyiv Polytechnic Institute, Ukraine

Accurate 3D reconstruction of dental structures is crucial for orthodontic assessment and surgical planning, yet traditional methods such as SIFT and ORB often struggle to capture fine details in complex dental textures. In this paper, we present a 3D reconstruction pipeline that combines KeyNetAffNetHardNet for feature detection and matching with Surface-Aligned Gaussian Splatting (SuGaR) for high-quality mesh reconstruction. By leveraging KeyNet for robust keypoint identification, AffNet for affine normalization, and HardNet for discriminative descriptors, our approach achieves a 25% reduction in computation time compared to advanced deep learning methods like LoFTR and DISK + LightGlue. To further optimize the 3D meshes, SuGaR aligns surface Gaussians to actual geometry, improving both structural accuracy and rendering fidelity. A new pipeline was evaluated using a set of high-resolution video frames from a single participant's dental panorama, achieving peak SSIM and PSNR scores of 0.9538 and 28.98, respectively — improvements of approximately 10% and 15% over conventional approaches. Our findings highlight how integrating learned feature matching and surface-aligned reconstruction can yield high-fidelity 3D dental models while maintaining efficiency, ultimately advancing diagnostic precision and treatment outcomes in dentistry.

3D Reconstruction, Keypoint Matching, Gaussian Splatting, Dental Imaging, Deep Learning, Computer Vision