DINE: Distance Is Not Enough -- Learning Global Deformation Priors for Robust Soft-Tissue Point Cloud Registration

2026-07-16

AI Summary

Non-rigid registration of soft-tissue point clouds under large deformations, noise, and outliers remains challenging due to unreliable correspondences and globally implausible deformations when relying solely on local distance metrics. This work proposes DINE, the first framework to incorporate learned global deformation statistical priors into point cloud registration. DINE employs a two-stage optimization strategy: it first pretrains a base model using Chamfer distance and then jointly optimizes the deformation field via maximum a posteriori estimation, integrating either PCA-based Gaussian or normalizing flow priors within the Robust-DefReg and DefTransNet backbones. Experiments demonstrate that DINE reduces Chamfer distance by 27–69% on the DeformedTissue dataset, improves robustness to outliers and Gaussian noise by 66% and 83%, respectively, and achieves performance gains of 59–79% under extreme deformation scenarios in SynBench.

Abstract

Non-rigid point cloud registration is central to soft-tissue shape analysis, but large deformations, noise, and outliers make correspondence estimation challenging. Most learning-based methods rely on local objectives such as Chamfer distance, which encourage point-wise proximity but do not constrain the global plausibility of the predicted deformation field. We address this limitation with DINE, a maximum a posteriori framework that augments distance-based registration with a learned statistical prior over displacement vector fields. DINE is applied to two registration backbones, Robust-DefReg and DefTransNet, using a two-stage strategy: a first-stage model is trained with Chamfer distance, its predicted deformation fields are used to estimate a prior, and the model is then refined with a combined distance and negative log-prior objective. We compare a full-field PCA Gaussian prior with a per-vector normalizing-flow prior. Experiments on DeformedTissue and SynBench show lower mean Chamfer distance under deformation and corruption. On DeformedTissue, DINE-PCA reduces Chamfer distance by approximately 27--69\% relative to the corresponding Stage-1 backbone across deformation levels, and improves robustness by up to 66\% for outliers and 83\% for Gaussian noise. On SynBench, improvements are modest at the smallest deformation levels and reach approximately 59--79\% from moderate to severe deformation. These results suggest that global deformation plausibility is an important constraint for reliable soft-tissue point cloud registration. (The code will be published soon.)

Problem

• non-rigid point cloud registration
• soft-tissue shape analysis
• global deformation plausibility
• correspondence estimation
• deformation field

Innovation

• non-rigid registration
• deformation prior
• point cloud
• statistical prior
• robustness