This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YANG et al.: SAC-COT FOR 3-D POINT CLOUD REGISTRATION 3
method. Section IV presents the deployed experiments to
validate the effectiveness of our method. Finally, Section V
concludes this article and presents future research directions.
II. R
ELATED WORK
This section first gives a brief review on 6-DOF pose estima-
tors, geometrics constraints, and deep learning techniques for
3-D registration. Then, since SAC-COT is a guided sampling-
based estimator, existing guided sampling approaches in the
RANSAC family is recapped.
A. 6-DOF Pose Estimators for 3-D Registration
For the problem of estimating a 6-DOF pose from cor-
respondences with outliers in the context of 3-D regis-
tration, RANSAC [15] and its variants remain the main-
stream solution. The classical RANSAC exhibits two main
limitations when faced with heavy outliers: low time effi-
ciency and limited accuracy. To overcome the two limita-
tions, Rusu et al. [10] presented a sample consensus-based
initial alignment (SAC-IA) method, which samples corre-
spondences spread out on the point cloud and leverages the
Huber penalty for hypothesis evaluation. To achieve more
reliable hypothesis evaluation, Yang et al. [7] proposed a point
cloud distance metric to assess pose hypotheses and intro-
duced the optimal sample consensus (OSAC) method. Both
SAC-IA and OSAC are three-point-based sampling methods
that exhibit a theoretical computational complexity of O(n
3
).
Some methods that sample fewer correspondences per itera-
tion have also been proposed. For instance, two-point-based
sample consensus with global constraint (2SAC-GC) [21] and
compatibility-guided sample consensus (CG-SAC) [1] addi-
tionally consider the normal information of keypoints during
the sampling process; globally constrained one-point-based
sample consensus (GC1SAC) [17] and one-point RANSAC
(1P-RANSAC) [16] additionally employ the LRF cue and only
sample one correspondence per iteration. Unfortunately, these
RANSAC methods still fail to achieve a good balance in terms
of accuracy, speed, and robustness to common nuisances (as
will be verified in Section IV).
In addition to RANSAC methods, Guo et al. [20] proposed a
clustering-based method that first generates pose hypotheses in
the rotation and translation spaces and then locates the cluster
centers in both spaces; Tombari and Di Stefano [18] and Buch
et al. [19] proposed two voting-based methods that serve pose
hypotheses as voters and perform voting in a Hough space. All
these methods are dependent on the normals or LRFs, which
have been demonstrated to be instable in the presence of noise,
partial overlap, and occlusions [13], [27], [28].
B. Geometric Constraints for 3-D Registration
Point-, plane-, and correspondence-level geometric con-
straints have been proposed to assist 3-D registration. For
point-level constraints, Aiger et al. [29] enforced coplanar
constraints to four-point sets in point clouds to achieve reg-
istration; Drost et al. [30] presented a point pair voting-
based method for 3-D registration, where distance and angle
constraints are employed to describe point pair features. For
plane-level constraints, Xu et al. [31] introduced a voxel-
based four-plane congruent set (V4PCS) method for pair-
wise coarse registration, which first extracts the planes for
point clouds and then leverages the plane-based geometric
constraints formed by four planes to achieve registration;
Chen et al. [32] presented a plane-based descriptor using dis-
tance and angle properties derived from planes for 3-D regis-
tration. For correspondence-level geometric constraints, most
of the existing methods focus on pairwise constraints. They
typically include the L
2
distance constraint [22], [33], normal
deviation angle constraint [21], [34], and LRF constraint [35].
Nonetheless, these pairwise constraints either suffer from
ambiguities or lack robustness to common nuisances. The
readers can refer to a related survey [36] for terrestrial laser
scanner point cloud registration for more details on geometrics
constraints proposed for 3-D registration.
SAC-COT leverages the correspondence-level geometric
constraints for registration, while it proposes a novel triplet
constraint that is validated to be far less ambiguous and more
robust than existing pairwise constraints.
C. Deep Learning for 3-D Registration
In addition to geometric-only methods, a few deep learning-
based point cloud registration methods have been pro-
posed recently. We review some typical examples here.
Aoki et al. [37] proposed the PointNetLK network for 3-D
registration in an iterative manner, where PointNet [38]
was employed for descriptor extraction combined with
a Lucas/Kanade-like optimization algorithm. Wang and
Solomon [39] presented deep closest point (DCP) that incor-
porates a point cloud embedding network, an attention-based
module to approximate combinatorial matching, and a differ-
entiable singular value decomposition (SVD) layer for 6-DOF
pose estimation. PRNet [40], as an extension to DCP, specifi-
cally focuses on partial-to-partial registration, which contains
a keypoint detection module and is able to establish partial
correspondences.
Deep learning techniques have demonstrated a great poten-
tial for 3-D registration, while usually, adequate labeled train-
ing data should be prepared. This may not be feasible in many
practical scenarios, and as such, we still focus geometric-only
methods.
D. Guided Sampling Approaches for RANSAC
As the key novelty of SAC-COT is its sampling method,
we herein review existing guided sampling methods for
RANSAC. In the 2-D domain, guided sampling approaches
preferentially sample correspondences with high confidence
scores [26], [41]. By contrast, guided sampling approaches
for RANSAC in the 3-D domain, as will be discussed in the
following, are more diverse.
For one-point-based sampling methods [16], [17], because
they have a linear time complexity, guided or random sampling
ways generate the same outcome as long as all correspon-
dences have been traversed. However, these one-point-based
sampling methods typically guide the sampling process based
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 14,2021 at 06:43:52 UTC from IEEE Xplore. Restrictions apply.