Input Data Processing Methods to Improve Point Cloud Completion Model for Dental Prosthesis

Teeth vary in size and shape among individuals, making the production of dental prosthetics with consistent quality challenging. Additionally, the quality of dental prosthetics is influenced by the manual skill level of dental technicians who create them. To improve this, recent studies have utilized generative adversarial networks (GANs) to create dental prosthetics. With 2D GAN models, a problem arises when converting 3D dental data into 2D data, resulting in the loss of spatial features. Point cloud completion models, which use 3D data as input, offer a partial solution to this problem. The performance of point cloud completion models varies depending on the representation and information contained in the input data. With this in mind, this study proposes three methods for processing input data to improve performance: changing the uniformity of the point cloud distribution, varying the number of sampled points, and including occlusal information with adjacent teeth. Experiments were conducted to compare and analyze performance changes based on these input data modifications. The results showed that there was up to a 30% performance difference in representative point cloud completion models, SnowFlakeNet and SeedFormer, based on CD (Chamfer Distance) Loss.