Memory-Guided Point Cloud Completion for Dental Reconstruction

TL;DR

The paper addresses unstable dental point-cloud completion under occlusions by introducing a retrieval-augmented prototype memory that debiases encoder features. A dual-encoder setup, a learnable memory bank of tooth-shaped prototypes, and a confidence-gated fusion with a folding-based decoder cooperate to provide structural priors that stabilize large missing regions and enhance fine geometry. The approach yields consistent improvements on Teeth3DS-derived data, achieving sharper cusps and better interproximal transitions, with ablations showing the crucial roles of the memory and dual encoders. The method offers a plug-and-play, label-free enhancement to existing completion backbones, enabling more faithful dental reconstructions in practical scanning scenarios.

Abstract

Partial dental point clouds often suffer from large missing regions caused by occlusion and limited scanning views, which bias encoder-only global features and force decoders to hallucinate structures. We propose a retrieval-augmented framework for tooth completion that integrates a prototype memory into standard encoder--decoder pipelines. After encoding a partial input into a global descriptor, the model retrieves the nearest manifold prototype from a learnable memory and fuses it with the query feature through confidence-gated weighting before decoding. The memory is optimized end-to-end and self-organizes into reusable tooth-shape prototypes without requiring tooth-position labels, thereby providing structural priors that stabilize missing-region inference and free decoder capacity for detail recovery. The module is plug-and-play and compatible with common completion backbones, while keeping the same training losses. Experiments on a self-processed Teeth3DS benchmark demonstrate consistent improvements in Chamfer Distance, with visualizations showing sharper cusps, ridges, and interproximal transitions. Our approach provides a simple yet effective way to exploit cross-sample regularities for more accurate and faithful dental point-cloud completion.

Figures (3)

• Figure 1: Overview of our retrieval-augmented completion framework. (a) Dual Feature Extractor: Two encoders with identical architecture but no weight sharing produce global descriptors for the partial input ($\mathcal{F}_{pi}$) and complete ground truth ($\mathcal{F}_{gt}$). (b) Retrieval-Augmented Memory Module: A learnable prototype bank $\{F_m^k\}_{k=1}^K$ is queried by features. During training, $\mathcal{F}_{gt}$ selects its nearest prototype $F_m^{j}$ to update the prototype memory ($\mathcal{L}_{\text{code}}$) and align distributions with the partial branch ($\mathcal{L}_{\text{align}}$). During inference, $\mathcal{F}_{pi}$ retrieves $F_m^{i}$; the prototype is confidence-gated and fused to yield the debiased global $\mathcal{F}'$. (c--d) We use a standard encoder backbone and a folding-based decoder to generate the completed point cloud from $\mathcal{F}'$.
• Figure 2: Qualitative dental completion results: $\mathcal{P}_{pi}$ (left), $\widehat{\mathcal{P}}$ (middle), $\mathcal{P}_{gt}$ (right).
• Figure 3: UMAP of encoder features and prototypes.