DentEdit3D is a system for language-driven editing of 3D intraoral scans. DentEdit3D is an experimental adaptation of ChangeIt3D into the dental domain. This project aims to experiment with the system and evaluate the results, focusing on a single geometry modification to a 3D shape, such as removing a specific tooth.
In order to train a whole system from scratch, one could use scripts in xtarag01/src/scripts:
create_dataset/create_dataset.shto create a dental contrastive dataset,train_pc_ae/train_pc_ae.shto train the PC-AE,train_latent_listener/train_latent_listener.shto train the Neural Listener,train_change_it_3d/train_change_it_3d.shto train the Shape Editor,evaluate_change_it_3d/evaluate_change_it_3d.shto evaluate the trained system on a test set.
To efficiently calculate CD and EMD on the GPU, specific CUDA PyTorch implementations of these metrics were used:
TARAGEĽ, Marián. Language-Guided 3D Anatomical Shape Editing. Brno, 2026. Master’s thesis. Brno University of Technology, Faculty of Information Technology. Supervisor Ing. Tibor Kubík
I would like to convey my gratitude to Ing. Tibor Kubík for his supervision. I also express my thanks to Prof. Yong-Jin Liu for providing the 3D Pre/Post-Orthodontic Dental Dataset. Computational resources were provided by the e-INFRA CZ project (ID:90254), supported by the Ministry of Education, Youth and Sports of the Czech Republic.