DualReg: Dual-Space Filtering and Reinforcement for Rigid Registration

Jiayi Li, Yuxin Yao, Qiuhang Lu, Juyong Zhang · 2025

Read DualReg: Dual-Space Filtering and Reinforcement for Rigid Registration by Jiayi Li, Yuxin Yao, Qiuhang Lu, Juyong Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-17034.txt.

Introduction

Jiayi Li, Yuxin Yao, Qiuhang Lu, and Juyong Zhang present DualReg, a 2025 method for aligning two three-dimensional scans of the same scene. The task, called rigid registration, is hard when the data are noisy, overlap only in part, and must be processed fast. Feature matching handles big shifts but lands imprecisely, while geometry matching is precise yet needs a good starting guess.

DualReg chains the two strengths. A featherweight one-point sampling filter, followed by a refinement pass, throws out unreliable feature matches. The survivors become anchor points, from which the method builds geometric stand-ins and solves for the shift with a tailored solver. On a standard driving data set, the pipeline ran about thirty-two times faster than a leading rival on a plain processor, with accuracy to match. The authors offer it for real-time mapping jobs.

Catalog record

CREATORJiayi Li, Yuxin Yao, Qiuhang Lu, Juyong Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17034

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DualReg: Dual-Space Filtering and Reinforcement for Rigid Registration Jiayi Li, Yuxin Yao, Qiuhang Lu, Juyong Zhang Abstract Rigid registration, aiming to estimate a rigid transformation to align source and target data, play a crucial role in applications such as SLAM and 3D reconstruction. However, noisy, partially overlapping data and the need for real-time processing pose major challenges for rigid registration. Considering that feature-based matching can handle large transformation differences but suffers from limited accuracy, while local geometry-based matching can achieve fine-grained local alignment but relies heavily on a good initial transformation, we propose a novel dual-space paradigm to fully leverage the strengths of both approaches. First, we introduce an efficient filtering mechanism that incorporates a computationally lightweight single-point RANSAC algorithm followed by a refinement module to eliminate unreliable feature-based correspondences. Subsequently, we treat filtered correspondences as anchor points, extract geometric proxies, and formulates an effective objective function with a tailored solver to estimate the transformation. Experiments verify our method’s effectiveness, as shown by achieving up to a 32x CPU-time speedup over MAC on KITTI with comparable accuracy. I Introduction Rigid registration is to estimate a rigid transformation for a source signal so that it can be transformed to be aligned with a target signal. It plays an important role in SLAM, robotics, 3D reconstruction, AR or VR, and 3D data processing. Since the signals used for registration are usually partially overlapping and often contain noise and outliers, moreover, some real-time applications, such as SLAM, place higher requirements on the efficiency. These make the fast and accurate registration algorithm is a challenging and critical task. Since correspondences are unknown, the Iterative Closest Point (ICP) algorithm [ 1 ] and its variants [ 2 , 3 , 4 , 5 ] establish correspondences by iteratively searching for closest points in Euclidean space, considering both spatial positions and geometric information of processing signals. Through alternating optimization of rigid transformations and correspondence establishment on transformed source and target signals, the process gradually converges to an optimal solution. These methods enable fine-grained local alignment but depend on the initial spatial positioning between source and target signals. When encountering substantial transformation discrepancies, they frequently converge to suboptimal local minima. Fig. 1: Registration Recall and Average Runtime (per-registration) on 3DMatch dataset with FPFH feature. Optimal methods require high recall and low runtime. Our method achieves the highest recall with lower time cost. Differently, some methods construct descriptors for processing signals and establish feature-based corresponding points globally based on the similarity of descriptors [ 6 , 7 , 8 ] . However, when the point cloud density is uneven or noisy, the quality will be greatly degraded. With the popularity of deep learning, many methods [ 8 , 9 , 10 , 11 ] design network structures to learn point cloud descriptors. Although these methods are dedicated to extracting better feature descriptors to establish better correspondences, it is inevitable that there will still be some outliers. Therefore, numerous outlier rejection methods are proposed to filter the inaccurate correspondences [ 10 , 12 , 9 , 13 , 14 , 15 , 16 ] . The traditional method is the RANSAC algorithm [ 17 ] , which iteratively generates hypotheses and estimates the model until it finds the subset with the highest inlier rate. But the RANSAC algorithm usually requires a large number of iterations, which may bring high computation costs. Therefore, some subsequent methods [ 18 , 19 , 20 , 21 , 22 ] have improved its performance. TCF [ 23 ] proposes one-point RANSAC and two-point RANSAC to greatly reduces the number of iterations, but it reduces the alignment quality. Recently, MAC [ 24 ] constructs the graph for the initial correspondences and generate hypotheses by searching for the maximal cliques. It greatly improves the registration accuracy, but takes more time. Follow-up works [ 25 , 26 ] are proposed to improve its speed or accuracy. However, a registration method that is both fast and precise remains essential. In this paper, we aim to develop an efficient and effective rigid registration algorithm capable of rapidly filtering out inaccurate correspondences from initial candidate correspondences and estimating precise rigid transformations. We observe that ICP-based methods yield corresponding points in local geometric space, which is fine-grained but depends on the initial transformation. Conversely, feature-space correspondences obtained through techniques such as [ 6 , 11 ] are transformation-invariant but often fail to achieve precise local alignment. To combine the strengths of both approaches, a feasible solution is to utilize feature-based methods to obtain a rough rigid transformation, which is then further refined by ICP-based methods to achieve high-precision registration. However, under low-overlap scenarios, ICP-based local registration algorithms often converge to suboptimal solutions due to erroneous nearest point correspondences. While robust norm-based ICP variants [ 27 , 5 ] can reduce sensitivity to outliers, their fundamental reliance on spatial proximity constraints limits their performance. To overcome these limitations, we propose a dual-space optimization framework that synergizes correspondences between global feature and local geometric spaces to achieve high-precision registration. To establish reliable correspondences within the feature space, we design an efficient correspondence filtering strategy that rapidly extracts high-confidence correspondences from raw matches generated using either hand-crafted or learned features. Our core innovation lies in the design of a computationally lightweight single-point RANSAC algorithm for preliminary correspondence filtering, followed by refinement using a three-point RANSAC algorithm. These high-confidence correspondences enable the construction of geometric proxy point sets with significantly improved overlap with the input point cloud. On these proxy point sets, we establish local geometry-based correspondences via a nearest neighbor search in Euclidean space. We then construct a dual-space optimization framework that synergistically integrates feature- and geometry-based correspondences and design an efficient iterative solver to compute the optimal rigid transformation. Experimental results demonstrate the effectiveness and efficiency of our approach. In summary, our main contributions include: • We design a dual-space optimization framework that synergistically integrates correspondences in feature and local geometric spaces to accurately estimate rigid transformations. • We propose an efficient progressive filtering mechanism for feature-based correspondences, which is achieved through fast filtering based on single-point RANSAC and improved accuracy of a probability-guided three-point RANSAC sampler. • Based on the filtered feature-based correspondences, we construct a geometric proxy point set and dynamically establish geometry-based correspondences, ultimately achieving accurate rigid registration. II Related Work In this section, we review the methods relevant to this paper as follows. Correspondence Prediction. Extracting accurate correspondences between source data and target data is crucial to achieving global alignment of objects or scenes. Various methods have been proposed to address this challenge. Some approaches select keypoints from point clouds and construct descriptors for them using hand-crafted approaches, such as FPFH [ 6 ] , SHOT [ 28 ] , etc., and establish correspondences between keypoints based on the similarity of their descriptors [ 7 ] . More methods adopt deep learning techniques to predict more accurate correspondences. LCGF [ 8 ] first extracts hand-crafted features and then maps them to a compact representation using a network. 3DMatch [ 10 ] extracts local 3D patches, converts them into a volumetric representation, and learns 3D ConvNet-based descriptors to establish the correspondences. [ 9 ] adopts multiple PointNets [ 29

Read the complete text