UniRiT: Towards Few-Shot Non-Rigid Point Cloud Registration

TL;DR

UniRiT adopts a two-step registration strategy that first aligns the centroids of the source and target point clouds and then refines the registration with non-rigid transformations, thereby significantly reducing the problem complexity.

Abstract

Non-rigid point cloud registration is a critical challenge in 3D scene understanding, particularly in surgical navigation. Although existing methods achieve excellent performance when trained on large-scale, high-quality datasets, these datasets are prohibitively expensive to collect and annotate, e.g., organ data in authentic medical scenarios. With insufficient training samples and data noise, existing methods degrade significantly since non-rigid patterns are more flexible and complicated than rigid ones, and the distributions across samples are more distinct, leading to higher difficulty in representation learning with few data. In this work, we aim to deal with this challenging few-shot non-rigid point cloud registration problem. Based on the observation that complex non-rigid transformation patterns can be decomposed into rigid and small non-rigid transformations, we propose a novel and effective framework, UniRiT. UniRiT adopts a two-step registration strategy that first aligns the centroids of the source and target point clouds and then refines the registration with non-rigid transformations, thereby significantly reducing the problem complexity. To validate the performance of UniRiT on real-world datasets, we introduce a new dataset, MedMatch3D, which consists of real human organs and exhibits high variability in sample distribution. We further establish a new challenging benchmark for few-shot non-rigid registration. Extensive empirical results demonstrate that UniRiT achieves state-of-the-art performance on MedMatch3D, improving the existing best approach by 94.22%.

Figures (14)

• Figure 1: UniRiT performs a rigid transformation phase between the source $\mathbf{P}_\mathcal{S}$ and target $\mathbf{P}_\mathcal{T}$ point clouds, where the features of both point clouds are extracted using MLPs. These features are then passed through a decoder composed of fully connected (FC) layers, which iteratively generates rotation and translation matrices over $n$ cycles. The transformed point cloud output from the rigid module is subsequently utilized along with the target point cloud to re-extract features. These features are concatenated with the coordinate information and then input into the decoder to generate a deformation matrix, which applied to $\mathbf{P}_\mathcal{S}'$, yields the final transformed point cloud $\hat{\mathbf{P}_\mathcal{S}}$.
• Figure 2: The ten types of the organs in MedMatch3D.
• Figure 3: In the comparison of visualization results for certain organs, the differences in pre-registration RMSE across different organ types are due to their different size and complexity. The blue point cloud represents the target point cloud, while the before image illustrates the discrepancy between the source and target point clouds before registration. In the method figure, the red point cloud indicates the transformed source point cloud.
• Figure 4: Seven randomly selected samples of the small bowel are shown. It can be observed that, during the acquisition of small bowel samples, issues such as incomplete structural scans and significant noise are present.
• Figure 5: The visualization results of Case B. For Case B, the non-rigid deformation magnitude is 15 mm, the rotation angle ranges from [0, 45°], and the translation range is [20, 30] mm.