3D Oral Modelling with Improved Vertex Distribution Using Matching-Based Learning

Researchers improved a deep learning framework for 3D oral reconstruction by introducing Hungarian matching and Repulsion Loss to achieve more uniform vertex distribution across predicted dental models. While numerical accuracy decreased from 77.49% to 68.02%, the trade-off eliminates vertex clustering in sparse regions, producing more clinically useful reconstructions from intraoral images.

This research addresses a fundamental challenge in 3D computer vision: balancing overall accuracy against practical distribution quality in predicted outputs. The original framework achieved impressive numerical accuracy but suffered from pathological clustering where predicted vertices concentrated in dense regions of training data, leaving sparse areas poorly represented. This phenomenon, common in point cloud prediction tasks, renders models less useful for real-world applications despite higher benchmark scores.

The improvement introduces Hungarian matching algorithms to enforce correspondences between predicted and ground truth vertices, combined with Repulsion Loss to encourage spatial separation. This represents a methodological shift from optimizing global metrics toward ensuring uniform coverage—a more nuanced approach to evaluating reconstruction quality. The deliberate accuracy trade-off (from 77.49% to 68.02%) reveals an important insight: raw numerical metrics can mask functional deficiencies in deep learning outputs.

For dental imaging and oral reconstruction applications, more evenly distributed vertices likely provide superior results for downstream tasks like treatment planning, prosthetic design, or surgical guidance. This work demonstrates that sometimes reducing benchmark scores yields better practical performance. The approach has implications beyond dentistry, potentially benefiting any 3D reconstruction task where uniform sampling matters more than peak accuracy metrics.

Future research should validate clinical utility through dentist evaluations and explore whether this distribution-focused approach generalizes to other point cloud prediction domains. The interplay between mathematical optimization and practical applicability highlighted here could influence how researchers design loss functions for computer vision tasks.

• →Hungarian matching and Repulsion Loss improve vertex distribution uniformity in 3D oral reconstruction despite lower numerical accuracy scores.
• →Vertex clustering in sparse regions was substantially eliminated, addressing a critical limitation of previous deep learning approaches.
• →The research demonstrates that higher benchmark accuracy doesn't always translate to better practical performance in dental reconstruction applications.
• →Loss function design profoundly impacts the quality and usability of predicted 3D point clouds beyond traditional metrics.
• →The framework uses MobileNetV2 and multi-head attention on ten fixed-angle intraoral images for real-time dental scanning applications.