Abstract
DentalSplat is a novel 3D reconstruction framework that generates high-fidelity dental models from as few as three sparse, unposed intra-oral photographs. By combining the "pose-free" initialization of DUSt3R with an optimized 3D Gaussian Splatting (3DGS) pipeline, it enables accurate remote orthodontic monitoring using only a smartphone.
Introduction
In orthodontics, tracking dental occlusion—how teeth move and fit together—traditionally requires expensive clinical equipment such as Cone Beam Computed Tomography (CBCT) or Intraoral Scanners (IOS). While telemedicine allows patients to capture photos at home, these images are typically "sparse" (e.g., one front view and two side views) and "unposed," lacking camera position data.
Existing 3D reconstruction methods like standard 3D Gaussian Splatting (3DGS) or Neural Radiance Fields (NeRF) often fail in these scenarios. They typically require dozens of overlapping images and precise camera calibration (e.g., via COLMAP). Dental photos also present unique challenges, including specular reflections from enamel and motion blur. DentalSplat addresses these issues, providing clinical-grade 3D visualization from minimal, consumer-grade input.
Methodology
The DentalSplat framework employs a multi-stage pipeline that eliminates the need for traditional Structure-from-Motion (SfM) dependencies.
1. Pose-Free Initialization
Instead of relying on COLMAP, the framework utilizes DUSt3R, a geometric foundation model. DUSt3R regresses a dense point cloud and estimates camera poses directly from image pairs, providing a robust starting point for reconstruction.
2. Scale-Adaptive Pruning (SAP)
Point clouds generated by DUSt3R are often excessively dense for efficient 3DGS training. The Scale-Adaptive Pruning (SAP) module analyzes the spatial distribution of these points and applies adaptive thresholds to prune the cloud. This retains critical geometry while significantly reducing computational overhead.
3. Optical Flow Constraints
To bridge the extreme angular gaps between sparse viewpoints, the authors incorporate a bidirectional optical flow loss. This acts as a geometric "glue," ensuring that 3D Gaussians remain consistent across different views even with minimal visual overlap.
4. Gradient-Weighted Optimization
To combat "Gradient Collision"—where conflicting signals from different pixels cause blurring—the authors apply an absolute operation to gradients. This ensures that the densification of Gaussians is guided by clear directional signals, resulting in sharper edges on tooth structures.
Experimental Evaluation
The researchers validated DentalSplat on two clinical datasets: a Video Test Dataset (195 cases) for rigorous evaluation and a 3-View Image Dataset (950 cases) representing the standard remote monitoring challenge.
Key Results
DentalSplat significantly outperformed state-of-the-art baselines like InstantSplat and CF-3DGS:
- Accuracy: On the 3-view dataset, DentalSplat achieved a PSNR of 34.50, compared to 32.78 for InstantSplat. Standard 3DGS failed to converge on such sparse data.
- Efficiency: The system completes a full 3D reconstruction in approximately 69 seconds on a consumer-grade RTX 4090 GPU.
- Visual Fidelity: Qualitative results demonstrate successful reconstruction of molar surfaces and gingival margins without the "melting" or "ghosting" artifacts common in other sparse-view methods.
Conclusion
DentalSplat demonstrates that while 3DGS is excellent at rendering, its performance in sparse settings is greatly enhanced by using foundation models like DUSt3R for initialization. By intelligently pruning initial point clouds and applying flow-based constraints, high-quality dental models can be generated from minimal input.
Future work may involve integrating domain-specific anatomical priors to improve robustness against poor lighting or extreme occlusion. Additionally, optimizing the pipeline for mobile-real-time execution could provide patients with instant feedback during the photo-taking process.
Summary updated