Dental3R: Geometry-Aware Pairing for Intraoral 3D Reconstruction from Sparse-View Photographs (2511.14315v1)

Abstract: Intraoral 3D reconstruction is fundamental to digital orthodontics, yet conventional methods like intraoral scanning are inaccessible for remote tele-orthodontics, which typically relies on sparse smartphone imagery. While 3D Gaussian Splatting (3DGS) shows promise for novel view synthesis, its application to the standard clinical triad of unposed anterior and bilateral buccal photographs is challenging. The large view baselines, inconsistent illumination, and specular surfaces common in intraoral settings can destabilize simultaneous pose and geometry estimation. Furthermore, sparse-view photometric supervision often induces a frequency bias, leading to over-smoothed reconstructions that lose critical diagnostic details. To address these limitations, we propose \textbf{Dental3R}, a pose-free, graph-guided pipeline for robust, high-fidelity reconstruction from sparse intraoral photographs. Our method first constructs a Geometry-Aware Pairing Strategy (GAPS) to intelligently select a compact subgraph of high-value image pairs. The GAPS focuses on correspondence matching, thereby improving the stability of the geometry initialization and reducing memory usage. Building on the recovered poses and point cloud, we train the 3DGS model with a wavelet-regularized objective. By enforcing band-limited fidelity using a discrete wavelet transform, our approach preserves fine enamel boundaries and interproximal edges while suppressing high-frequency artifacts. We validate our approach on a large-scale dataset of 950 clinical cases and an additional video-based test set of 195 cases. Experimental results demonstrate that Dental3R effectively handles sparse, unposed inputs and achieves superior novel view synthesis quality for dental occlusion visualization, outperforming state-of-the-art methods.

• The paper introduces a geometry-aware pairing strategy (GAPS) that optimizes image selection for accurate intraoral 3D reconstruction from sparse-view photographs.
• It employs a wavelet-regularized 3D Gaussian Splatting method to preserve high-frequency dental details while mitigating frequency bias in sparse data.
• It validates superior performance on clinical datasets, outperforming existing methods in metrics like PSNR, SSIM, and LPIPS, and enabling effective tele-orthodontics.

Dental3R: Geometry-Aware Pairing for Intraoral 3D Reconstruction from Sparse-View Photographs

Introduction

Dental3R introduces a novel approach to intraoral 3D reconstruction from sparse and unposed photographs, targeting the limitations of conventional clinical imaging workflows for digital orthodontics. Traditional approaches, such as intraoral scanning (IOS) or CBCT, are not viable for remote or tele-orthodontic scenarios, which increasingly rely on user-acquired, non-standardized, and sparse-view photographs. Recent neural rendering advancements, notably 3D Gaussian Splatting (3DGS), have advanced photorealistic synthesis from images, but these methods deteriorate in challenging settings characterized by wide baselines, complex specularities, and scarce viewpoints.

Dental3R addresses these challenges via a tightly integrated pipeline that couples a geometry-aware image pairing strategy (GAPS) with wavelet-regularized 3DGS optimization. This produces dense, geometrically faithful reconstructions even under extreme data sparsity, and empirically yields state-of-the-art view synthesis results in large-scale, real-world clinical datasets.

Methodology Overview

Dental3R comprises several distinctive components:

• Geometry-Aware Pairing Strategy (GAPS): A graph-guided framework for selecting image pairs robustly, facilitating accurate pose/geometry estimation without a priori camera information.
• Stereo-dense Point Cloud Initialization: Utilizing DUSt3R, geometry and camera poses are simultaneously regressed from the optimal viewpair subgraph.
• 3DGS Optimization with Wavelet Regularization: The subsequent 3D Gaussian Splatting step leverages wavelet-domain constraints to mitigate frequency bias—preserving crucial dental edge and textural fidelity for clinical usage.

The end-to-end pipeline is illustrated following the initial image pair selection, which underpins the stability and efficiency of subsequent stages:

Figure 1: Dental3R pipeline. GAPS generates high-value image pairs for robust DUSt3R-based geometry initialization, enabling dense 3DGS optimization with wavelet constraints.

Geometry-Aware Pairing and Graph Construction

GAPS formalizes the image connectivity graph to avoid the practical failures of both “complete” (quadratically scaling) and overly-sparse or star-like pairings. Given $n$ views arranged on a cyclic graph, GAPS parameterizes a set of multi-scale local and global chords via offsets $\mathcal{K}$, inducing candidate edges connecting view indices $(i, (i+k)\bmod n)$ for each $k \in \mathcal{K}$.

An importance score $w(i,j)$ is computed for each chord, reflecting geometric overlap and positional range (local, medium, long). A monotonic decay maps the circular graph distance $d(i,j)$ to a surrogate of expected view overlap, ensuring the edge set remains informative yet sparse. A degree-bounded matching is performed to maximize total importance while capping vertex degrees, yielding a compact, high-value subgraph.

This GAPS subgraph guides the DUSt3R pointmap regressor, which outputs globally aligned dense point clouds and robust estimates of relative camera poses—pivotal for initializing 3DGS, especially under wide baseline and specular intraoral conditions.

3D Gaussian Splatting with Wavelet Regularization

Dental3R builds on the 3DGS paradigm, representing the scene with explicit, anisotropic Gaussians whose attributes (position, covariance, color, opacity) are optimized via differentiable rasterization. However, with sparse-view photometric supervision, 3DGS typically induces frequency bias, oversmoothing critical dental details such as enamel boundaries and occlusal contacts.

To combat this, Dental3R augments the standard photometric loss with a two-level discrete wavelet transform (DWT)–based regularization term:

$$\mathcal{L}_\text{wavelet} = \sum_{x \in \{LL, LH, HL, HH\}} \lambda_x \| W_x(I_\text{gt}) - W_x(\hat{I}) \|_2^2$$

Here, $W_x$ denotes the wavelet sub-band extraction, emphasizing reconstruction fidelity across frequency bands—most critically for the preservation of high-frequency edges and textural variations. This regularization is essential for dental applications, where diagnostic accuracy depends on fine structural delineation.

Experimental Evaluation

The empirical evaluation leverages a comprehensive, clinically acquired intraoral dataset: 950 cases (standard three-view) and 195 video-based cases with 24 frames each. The framework is subjected to varying input sparsity regimes (3–12 views) and benchmarked against leading approaches: 3DGS, InstantSplat, and CF-3DGS.

Quantitative metrics include PSNR, SSIM, and LPIPS across withheld test views. Notably:

• 3DGS fails to converge under the most sparse conditions (as few as three images) and produces inferior quality even with dense input.
• Dental3R consistently surpasses all baselines in all metrics, with significant PSNR and SSIM gains and lower LPIPS, robust to input sparsity.
• Rendering fidelity and geometric accuracy are striking under the protocol’s target regime (three-view and six-view intraoral images), outperforming all contemporary graph/pairing strategies and achieving strong memory efficiency.

Qualitative visualizations reinforce these results:

Figure 2: Novel view synthesis; Dental3R delivers sharper teeth boundaries and reduced geometric or texture artifacts compared with 3DGS, InstantSplat, and CF-3DGS, especially with limited input views.

An ablation distills the contribution of GAPS and the wavelet constraint, showing that the GAPS strategy achieves performance competitive in quality with exhaustive pairings but at a fraction of computational and memory cost, while wavelet regularization is necessary for preserving structurally critical frequency details.

Figure 3: Novel view synthesis under alternative pair strategies; GAPS outperforms "oneref" and "cosine" graphs and approaches the exhaustive (“complete”) strategy—preserving diagnostic textural details while optimizing computational use.

Implications and Future Developments

Practically, Dental3R has immediate translational value for tele-orthodontics, enabling clinical-grade 3D models from minimal user-contributed images, supporting remote assessment, appliance design, and treatment tracking without specialized equipment or operator training.

Theoretically, the pipeline exemplifies a successful integration of structure-aware input selection and frequency-regularized neural rendering under severe data sparsity. The GAPS selection methodology and wavelet-domain loss offer promising directions for other medical imaging modalities, where data collection is subject to physical and logistical constraints, and fine-grained structural recovery is non-negotiable.

Potential future research includes:

• Extension to deformable tissue or dynamic intraoral sequences, leveraging recent advances in time-resolved 3DGS and multi-session fusion.
• Unsupervised or self-supervised rendering objectives exploiting synthetic or in silico data for clinical generalization.
• Downstream diagnostic and monitoring models that incorporate uncertainty or per-vertex quality scores derived from Dental3R’s confidence maps.

Conclusion

Dental3R systematically resolves the major bottlenecks in sparse-view intraoral photogrammetry. By unifying a geometry-aware pair selection (GAPS), robust pose-free point cloud initialization (DUSt3R), and frequency-preserving 3DGS optimization (wavelet constraints), it achieves precise, artifact-resistant dental reconstructions. The resulting models facilitate reliable clinical assessment from user-acquired photographs and substantially outperform all existing sparse-view baselines, positioning Dental3R as a foundational step towards scalable, AI-driven tele-orthodontics and 3D dental modeling (2511.14315).