Graph-matching-based correspondence search for nonrigid point cloud registration☆
Introduction
Point cloud registration is a key process for capturing full shapes of 3D objects. Capturing the point cloud of a 3D object is a fundamental process for computer vision and robotics problems such as 3D localization, 3D scene/object recognition, 3D reconstruction in augmented/virtual reality and generating free-viewpoint videos. However, capturing a high-quality 3D object in a single shot is challenging due to missing points occluded by other parts or sparse points in parts captured from distant surfaces. These are inevitable shortcomings we face in acquiring 3D data. Registration of simultaneously captured multi-view point clouds can complement missing and low-resolution regions, however, this is not enough for capturing the hidden sides of an object, such as an armpit or a thigh of a human body. On the other hand, when an object is deformable, nonrigid registration can supplement these by combining point clouds of different frames representing different poses of the object. For example, a hidden or missing human body part of one frame can be complemented by another frame in which the corresponding part is exposed to the camera.
Nonrigid registration can transform another point cloud of the same deformable object into the reference point cloud. Nonrigid registration algorithms generally consist of two procedures, one is finding correspondences and the other is updating parameters in the way that reduces some metric between the correspondence pairs. The difference of nonrigid registration from rigid registration is that nonrigid registration finds transformation parameters for each point to describe the motion to fit a source frame to a target frame whereas rigid registration finds one set of transformation parameters. Hence, for nonrigid registration, it is important to find a correct match for every point to determine where to transform it. In order to acquire a high-quality 3D shape from a deformable object, there are two requirements: A good registration algorithm that can find correct matches between two point clouds of an object that has large motions, and an ability to process two massive point clouds containing hundreds of thousands of points. However, satsifying both conditions is difficult for nonrigid registration.
The nonrigid iterative closest point (nonrigid ICP) method, the most popular nonrigid registration technique, finds correspondences as the closest points in every iteration hoping to find correct matches in the end. However, nearest search based on point position are sensitive to noises. Various feature extraction methods like Shape Contexts (SC) (Belongie et al., 2002) and Point Feature Histogram (PFH) (Rusu et al., 2008a, Rusu et al., 2008b) are developed to improve correspondence search by characterizing local surface and shape of each points. Further, Chui and Rangarajan (2003) relax binary correspondence condition of ICP to soft-assignment. On the other hand, methods to find robust transformations from erroneous matching are developed. Ma et al. (2014) remove outliers from a putative correspondence set containing erroneous matching and Ma et al., 2015, Ma et al., 2013 estimate transformation parameters from erroneous matching using robust estimator. However, the performance degrades when an underlying motion of two point sets is large and often computing features for massive point sets are inexecutable due to memory limit. Another main stream of nonrigid registration is to formulate registration problem as a probability density estimation problem modeled as Gaussian Mixture Models (Jian and Vemuri, 2010, Wang et al., 2015, Wang et al., 2014, Myronenko and Song, 2010). These approaches are proposed to handle noises, missing, and outliers, however, they are prone to errors when two point clouds under registration are related to large motion.
In this work, we propose a novel correspondence matching and initial alignment methods for nonrigid registration based on graph matching. Graph matching finds correspondences of vertices of two graphs by maximizing similarities between vertices and edges of the two graphs. This makes the matching process consider neighborhood connectivity, which contains both the local and global information. In order to incorporate graph matching in nonrigid registration, we first divide point clouds into a number groups and apply graph matching to find groupwise correspondences. In addition, we relax one-to-one matching problem to many-to-one matching problem to improve registration performance. As a consequence, our method can effectively match semantically identical human body parts of two frames even when the motions are large. Furthermore, the search based on a small number of point groups reduces memory usage compared to other nonrigid registration methods based on dense features, which makes our method favorable in handling large-scale point clouds. Using the matching results obtained from graph matching, we compute groupwise transformation parameters to find initial nonrigid alignment. These initial alignments are coarse but they provide great clues for the following refinement step. We evaluate our methods on the Lincoln human motion dataset (Collet et al., 2015) and the TOSCA animal dataset (Bronstein et al., 2008). In short, the contribution of the paper can be summarized as follows:
- •
We propose a novel graph-matching-based correspondence search method for nonrigid registration, which provide coarsely aligned point cloud even when the underlying motion is large.
- •
We propose many-to-one matching problem for nonrigid registration and its solver to enhance robustness of matching and empirically show that many-to-one matching preserves more local structures.
- •
We propose an initial alignment method which refines the rough registration results produced by graph matching procedure.
- •
We propose a nonrigid registration framework which can handle large-scale point clouds for high-quality 3D reconstruction.
The rest of the paper is organized as follows. Short review of a literature are introduced in 2. The proposed graph-matching-based nonrigid registration method is explained module by module in Section 3. Qualitative and quantitative experiments on human body and animal body datasets are conducted in Section 4, followed by conclusion and future work in Section 5.
Section snippets
Related work
In this section we briefly review the background material our work is based on.
Proposed method
The proposed graph-matching-based nonrigid registration method consists of three main modules: First, the graph matching module divides point clouds of two frames, i.e., source and target, into groups, and finds the correspondences between them under the assumption that the motion of each group can be approximated as a rigid transform. This module inevitably incurs some errors due to some incorrect matches or the rigid approximation, so the subsequent modules refine the transformed source point
Experiments
In this section, we evaluate the proposed method on the Lincoln (Collet et al., 2015) and the TOSCA (Bronstein et al., 2008) datasets. Lincoln consists of multiple frames of human body scans. We use the upper part of the body for registration as its motion is large whereas the lower part hardly moves. TOSCA consists of scans of animals and humans in various poses. Originally, all the datasets are in the form of mesh structures, consisting of vertices and faces. We use the vertices of a mesh as
Conclusions
In this paper, we proposed a novel nonrigid registration algorithm which registers two different point clouds capturing different poses of a single deformable object, by finding corresponding parts between the point clouds using graph matching. In particular, our algorithm has advantages when registering two frames with large displacements. We validated our algorithm extensively on the Lincoln and TOSCA datasets and showed that our algorithm is indeed successful in registering point clouds with
CRediT authorship contribution statement
Seunggyu Chang: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing - original draft. Chanho Ahn: Methodology, Software, Formal analysis. Minsik Lee: Conceptualization, Methodology, Writing - review & editing, Funding acquisition. Songhwai Oh: Conceptualization, Methodology, Writing - review & editing, Supervision, Funding acquisition.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by ‘The Cross-Ministry Giga KOREA Project’ grant funded by the Korea government (MSIT), Republic of Korea (No. GK20P0300, Real-time 4D reconstruction of dynamic objects for ultra-realistic service).
References (38)
- et al.
A new point matching algorithm for non-rigid registration
Comput. Vis. Image Underst.
(2003) - et al.
Performance of simulated annealing-based heuristic for the unconstrained binary quadratic programming problem
European J. Oper. Res.
(2001) - et al.
A survey for the quadratic assignment problem
European J. Oper. Res.
(2007) - et al.
A robust non-rigid point set registration method based on asymmetric gaussian representation
Comput. Vis. Image Underst.
(2015) - et al.
Articulated body deformation from range scan data
ACM Trans. Graph.
(2002) - et al.
Optimal step nonrigid ICP algorithms for surface registration
- et al.
The correlated correspondence algorithm for unsupervised registration of nonrigid surfaces
- et al.
Shape matching and object recognition using shape contexts
IEEE Trans. Pattern Anal. Mach. Intell.
(2002) - et al.
Method for registration of 3-D shapes
- et al.
Distributed optimization and statistical learning via the alternating direction method of multipliers
Found. Trends Mach. Learn.
(2011)
Numerical Geometry of Non-Rigid Shapes
Reweighted random walks for graph matching
High-quality streamable free-viewpoint video
ACM Trans. Graph.
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
Commun. ACM
A guide to the theory of NP-completeness
Reassembling fractured objects by geometric matching
ACM Trans. Graph.
Robust Statistics
Robust point set registration using gaussian mixture models
IEEE Trans. Pattern Anal. Mach. Intell.
Modeling deformable objects from a single depth camera
Cited by (15)
Improving sparse graph attention for feature matching by informative keypoints exploration
2023, Computer Vision and Image UnderstandingRobust real-world point cloud registration by inlier detection
2022, Computer Vision and Image UnderstandingCitation Excerpt :The non-overlapping outliers are defined for those non-overlapping points in the partially overlapping point clouds, whilst the overlapping outliers are defined as points contaminated by density differences or sensor noise in the overlapping regions. As a long-standing problem in many fields, there has been many efforts to resolve the point cloud registration (Bouaziz et al., 2013; Yang et al., 2015; Huang et al., 2017a; Mellado et al., 2014; Zhou et al., 2016; Aoki et al., 2019; Huang et al., 2020; Yuan et al., 2020; Pribanić et al., 2019; Ma et al., 2020; Huang et al., 2017b, 2016; Chang et al., 2020; Al-Nuaimi et al., 2017; Bai et al., 2021; Huang et al., 2021, 2019; Du et al., 2019). The existing correspondence-based registration methods can be divided into two categories: optimization-based and learning-based methods.
Non-rigid point set registration based on Gaussian mixture model with integrated feature divergence
2024, Robotic Intelligence and AutomationNon-rigid point set registration: recent trends and challenges
2023, Artificial Intelligence ReviewPoint Cloud Registration Method Based on Improved Covariance Matrix Descriptor
2023, Xitong Fangzhen Xuebao / Journal of System SimulationVirtual Restoration of Ancient Wooden Ships Through Non-rigid 3D Shape Assembly with Ruled-Surface FFD
2023, International Journal of Computer Vision
- ☆
No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.cviu.2019.102899.