Diff-PCR: Diffusion-Based Correspondence Searching in Doubly Stochastic Matrix Space for Point Cloud Registration

TL;DR

This paper introduces Diff-PCR, a diffusion-based framework for point cloud registration that explicitly refines correspondence matrices through iterative denoising, leading to more accurate and optimal solutions.

Contribution

It proposes a novel diffusion model approach to explicitly refine correspondence matrices, improving over existing one-shot methods in point cloud registration.

Findings

Effective in 3DMatch/3DLoMatch and 4DMatch/4DLoMatch datasets.

Outperforms existing methods in correspondence accuracy.

Uses a lightweight denoising module with accelerated sampling.

Abstract

Efficiently identifying accurate correspondences between point clouds is crucial for both rigid and non-rigid point cloud registration. Existing methods usually rely on geometric or semantic feature embeddings to establish correspondences and then estimate transformations or flow fields. Recently, several state-of-the-art methods have adopted RAFT-like iterative updates to refine solutions. However, these methods still have two major limitations. First, their iterative refinement mechanism lacks transparency, and the update trajectory is largely fixed once the refinement starts, which may lead to suboptimal solutions. Second, they overlook the importance of explicitly refining the correspondence matrix before solving for transformations or flow fields. Most existing approaches compute candidate correspondences in feature space and project the resulting matching matrix only once by using Sinkhorn or dual-softmax normalization. Such a one-shot projection can be far from the globally optimal solution, and these methods usually do not model the distribution of the target matching matrix. In this paper, we propose a novel framework that exploits a denoising diffusion model to predict a search gradient for the optimal matching matrix in doubly stochastic matrix space. Specifically, the diffusion model learns a denoising direction, and the reverse denoising process iteratively searches for improved solutions along this learned direction, which approximates the maximum-likelihood direction of the target matching matrix. To improve efficiency, we design a lightweight denoising module and adopt the accelerated sampling strategy of the Denoising Diffusion Implicit Model (DDIM)\cite{song2020denoising}. Experimental results on 3DMatch/3DLoMatch and 4DMatch/4DLoMatch demonstrate the effectiveness of the proposed framework.