DC-Reg: Global Optimal Registration

• The paper's main contribution is formulating registration as simultaneous transformation and correspondence estimation, enabling robustness to partial overlaps and severe misalignment.
• It employs a branch-and-bound approach enhanced by a holistic concave underestimator from Difference-of-Convex programming, yielding tighter bounds and faster global search.
• The method efficiently solves cardinality-constrained linear assignment problems in low-dimensional transformation spaces, demonstrating competitive performance on both 2D and 3D benchmarks.

Searching arXiv for the target paper and closely related registration work for citations. DC-Reg is a globally optimal point cloud registration framework for settings in which two point sets may only partially overlap, may be severely misaligned, and may contain many outliers. It formulates registration as the simultaneous estimation of transformation parameters and correspondences, optimizing jointly over a transformation vector $\boldsymbol{\theta}$ and an assignment matrix $\mathbf{P}$, rather than treating correspondence as a byproduct of pose estimation. The method’s defining contribution is a branch-and-bound procedure equipped with a “holistic concave underestimator” derived from Difference-of-Convex programming, designed to tighten lower bounds for the coupled transformation-assignment objective and thereby accelerate globally optimal search (Lian et al., 26 Mar 2026).

1. Problem formulation and scope

DC-Reg addresses the registration problem through the coupled objective

$$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$

where $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$ lies in a search box

$$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$

and $\mathbf{P}=\{p_{ij}\}$ belongs to a feasible assignment set $\mathcal P$ (Lian et al., 26 Mar 2026). The corresponding marginalized objective is

$$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$

In implementation, the feasible correspondence set is

$$\mathcal{P} = \left\{\mathbf{P} \in \{0,1\}^{N \times M} \;\middle|\; \mathbf{P}\mathbf{1} \le \mathbf{1},\; \mathbf{1}^\top \mathbf{P} \le \mathbf{1}^\top,\; \sum p_{ij} = n_p \right\}.$$

This makes $\mathbf{P}$ a binary partial permutation matrix: each source point matches at most one target point, each target matches at most one source point, and exactly $\mathbf{P}$0 matches are selected (Lian et al., 26 Mar 2026). The mechanism for handling outliers and partial overlap is therefore combinatorial rather than penalty-based: unmatched points are excluded because only $\mathbf{P}$1 pairs are chosen.

The paper’s central viewpoint is that simultaneous transformation and correspondence estimation is attractive because correspondence flexibility makes the formulation naturally tolerant to partial overlap and some forms of nonrigid disturbance. The method does not directly model nonrigid deformation with deformation parameters; rather, it gains robustness to moderate nonrigid effects because the assignment can select a subset of plausible matches rather than forcing all points to fit a single rigid model (Lian et al., 26 Mar 2026). This suggests that DC-Reg occupies an intermediate position between classical rigid registration and explicitly nonrigid registration: it remains a rigid or similarity optimizer at the transformation level, but with a partial-assignment model that can absorb deviations from exact rigidity.

The paper studies this formulation for a 4-parameter 2D similarity transform and for 3D rigid registration approached through a 3D translation search followed by separate rotation recovery (Lian et al., 26 Mar 2026). A plausible implication is that the framework is most effective when the transformation search space remains low-dimensional, because its lower-bound construction requires vertex evaluation over transformation boxes.

2. Transformation models and registration energies

For 2D similarity registration, the transformation is parameterized by $\mathbf{P}$2, with $\mathbf{P}$3 the translation and

$\mathbf{P}$4

encoding scale $\mathbf{P}$5 and rotation angle $\mathbf{P}$6 (Lian et al., 26 Mar 2026). The registration energy is

$\mathbf{P}$7

with

$\mathbf{P}$8

Here $\mathbf{P}$9 gives the transformed source point under the 2D similarity parameterization, and the cost is purely spatial least squares (Lian et al., 26 Mar 2026).

For 3D registration, DC-Reg uses a rotation-invariant feature based on point norms to search only over translation $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$0: $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$1 Because

$$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$2

for any rotation $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$3, the translation search becomes independent of the unknown rotation (Lian et al., 26 Mar 2026). After $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$4 is found globally, the remaining rotation is recovered by an external rotation search method from the authors’ prior work. The paper presents this as an efficient instantiation of the framework rather than a direct 6D simultaneous rigid search (Lian et al., 26 Mar 2026).

This decomposition is significant because it defines the precise sense in which the 3D pipeline is globally optimal. The translation-assignment subproblem in the rotation-invariant feature space is solved globally; the full rigid registration is then completed by a separate rotation solver (Lian et al., 26 Mar 2026). It is therefore inaccurate to describe the 3D version as a direct globally optimal 6D rigid-registration solver. A more precise characterization is a globally optimal translation-assignment framework embedded in a two-stage rigid-registration pipeline.

3. Difference-of-Convex lower bounding and branch-and-bound

The key technical contribution is the Difference-of-Convex perspective used to derive a tighter lower bound inside branch-and-bound than earlier term-wise envelope constructions. The paper assumes each pairwise cost admits a DC decomposition

$$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$5

where $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$6 is convex and $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$7 is concave (Lian et al., 26 Mar 2026). Given a search box $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$8 and a linearization point $$E(\mathbf{P}, \boldsymbol{\theta}) = \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}),$$9, the convex part is under-estimated by its first-order affine lower bound,

$\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$0

yielding the concave underestimator

$\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$1

The associated auxiliary function is

$\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$2

Because each $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$3 is concave, the paper states that $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$4, as the pointwise minimum of concave functions, is also concave. Therefore minimizing $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$5 over a box $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$6 reduces to vertex evaluation: $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$7 where $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$8 is the set of the $\boldsymbol{\theta}\in\mathbb{R}^{n_\theta}$9 vertices of $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$0 (Lian et al., 26 Mar 2026).

The “holistic concave underestimator” is defined in opposition to earlier methods such as RPM-HTB or HBSP, which derive lower bounds by relaxing each bilinear or trilinear term independently, often using McCormick-type convex envelopes (Lian et al., 26 Mar 2026). DC-Reg instead first views each entire pairwise cost $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$1 as a DC function, constructs one underestimator $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$2 for that whole pairwise term, and then keeps the resulting transformation-dependent costs coupled with $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$3 inside

$$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$4

The paper’s claim is that this better captures the joint structure of pose and correspondence, leading to tighter bounds and faster pruning (Lian et al., 26 Mar 2026).

The branch-and-bound procedure partitions only the transformation space, not the correspondence variables. At each iteration it computes a lower bound $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$5 and candidate solution $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$6 for each newly created box; updates the global upper bound by evaluating the original objective at the best candidate among previous and new ones; prunes boxes satisfying

$$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$7

terminates when no boxes remain; otherwise selects the box with smallest lower bound and bisects its longest edge to create two child boxes (Lian et al., 26 Mar 2026). The paper explicitly states that if all remaining boxes are pruned, then the current best $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$8 is the global $$\mathcal{M} = [\underline{\boldsymbol{\theta}}, \overline{\boldsymbol{\theta}}],$$9-minimum solution (Lian et al., 26 Mar 2026).

4. Linear assignment reduction and concrete instantiations

Once $\mathbf{P}=\{p_{ij}\}$0 is fixed to a box vertex $\mathbf{P}=\{p_{ij}\}$1, the lower-bound evaluation becomes

$\mathbf{P}=\{p_{ij}\}$2

which is an $\mathbf{P}=\{p_{ij}\}$3-cardinality Linear Assignment Problem with cost matrix

$\mathbf{P}=\{p_{ij}\}$4

(Lian et al., 26 Mar 2026). In implementation, the paper uses the Jonker–Volgenant algorithm for these LAPs. Because the search dimension is low in the studied cases, each branch-and-bound node requires only $\mathbf{P}=\{p_{ij}\}$5 LAP evaluations for 2D similarity and $\mathbf{P}=\{p_{ij}\}$6 for 3D translation (Lian et al., 26 Mar 2026).

For the 2D similarity case, each pairwise cost

$\mathbf{P}=\{p_{ij}\}$7

is convex in $\mathbf{P}=\{p_{ij}\}$8, so the DC decomposition is trivial: $\mathbf{P}=\{p_{ij}\}$9 The underestimator becomes the first-order Taylor lower bound

$\mathcal P$0

with

$\mathcal P$1

This yields an especially clean implementation in which the bound reduces to evaluating affine costs at box vertices and solving cardinality-constrained assignments (Lian et al., 26 Mar 2026).

For the 3D translation search, the pairwise cost

$\mathcal P$2

admits the explicit DC decomposition

$\mathcal P$3

$\mathcal P$4

Linearizing the convex part at $\mathcal P$5 gives

$\mathcal P$6

with

$\mathcal P$7

and thus

$\mathcal P$8

Since translation is 3D, only eight vertices are checked per box, which the paper identifies as the main source of efficiency in the 3D variant (Lian et al., 26 Mar 2026).

The dominant cost per branch-and-bound node is therefore

$\mathcal P$9

with practical behavior determined jointly by search dimension and bound tightness (Lian et al., 26 Mar 2026). The authors further note that a direct 3D affine version would be impractical because $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$0 vertex evaluations per box would be required (Lian et al., 26 Mar 2026).

5. Robustness model, guarantees, and limitations

DC-Reg handles missing correspondences and outliers through the assignment constraints and the fixed cardinality $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$1. Because only $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$2 matches are selected, surplus points in either set are treated as outliers; the row and column inequality constraints impose one-to-one matching among selected pairs; and the cardinality controls how much overlap is presumed (Lian et al., 26 Mar 2026). There are no dummy nodes, slack variables, truncation penalties, or explicit outlier costs in the formulation.

This design yields a combinatorially clean partial-assignment model compatible with LAP solvers, but it also introduces a modeling dependence on $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$3. The paper explicitly studies sensitivity to this choice by using $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$4 and $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$5 of the ground-truth inlier count in experiments (Lian et al., 26 Mar 2026). A plausible implication is that the method is most attractive when a reasonable inlier-count prior is available, even if that prior is somewhat underestimated.

The theoretical statements are property-based rather than theorem-heavy. The central validity arguments are that $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$6 is a valid underestimator because a convex function dominates its tangent affine approximation; $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$7 is concave as a pointwise minimum of concave functions; and minimizing a concave function over a box is achieved at a vertex, making $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$8 valid (Lian et al., 26 Mar 2026). Combined with a standard branch-and-bound scheme over a compact initial box and exact lower bounds for each box, the method inherits the usual global $$E(\boldsymbol{\theta}) = \min_{\mathbf{P}\in\mathcal P} \sum_{i,j} p_{ij} E_{ij}(\boldsymbol{\theta}).$$9-optimality guarantee of deterministic branch-and-bound (Lian et al., 26 Mar 2026).

The authors also argue qualitatively that the bound is tighter than previous term-wise relaxations because it preserves variable interactions that McCormick-style envelopes ignore, but they do not provide a formal dominance theorem over prior bounds in the text shown (Lian et al., 26 Mar 2026). This is an important interpretive boundary: the empirical acceleration claim is strong, but the comparative tightness claim remains qualitative in the paper’s formal development.

The principal limitations identified in the paper are twofold. First, computational overhead still grows with branching and transformation dimension (Lian et al., 26 Mar 2026). Second, the framework does not directly model richer nonrigid deformations (Lian et al., 26 Mar 2026). The authors suggest future work on extending the DC paradigm to nonrigid models and accelerating branch-and-bound with GPU-parallel pruning (Lian et al., 26 Mar 2026).

6. Empirical evaluation and relation to prior global registration methods

DC-Reg is evaluated on synthetic and real 2D tests, synthetic 3D tests, and the 3DMatch benchmark (Lian et al., 26 Mar 2026). In 2D synthetic experiments on “fish” and “character” shapes under deformation, positional noise, mixed outliers, separate outliers, and occlusion plus outliers, baselines are RPM-HTB, RPM-PA, and RPM-CAV (Lian et al., 26 Mar 2026). The evaluation metric is RMSE between transformed source inliers and corresponding target inliers. The paper reports that DC-Reg is substantially more stable under deformation and positional noise, and especially strong under mixed and separated outliers, where it maintains near-zero error while baselines degrade; even when $$\mathcal{P} = \left\{\mathbf{P} \in \{0,1\}^{N \times M} \;\middle|\; \mathbf{P}\mathbf{1} \le \mathbf{1},\; \mathbf{1}^\top \mathbf{P} \le \mathbf{1}^\top,\; \sum p_{ij} = n_p \right\}.$$0 is underestimated to half the true inlier count, it remains competitive and often superior (Lian et al., 26 Mar 2026).

On 2D real-image experiments using Canny edge point sets from Caltech-256 and VOC2007 images, with source sets rotated by $$\mathcal{P} = \left\{\mathbf{P} \in \{0,1\}^{N \times M} \;\middle|\; \mathbf{P}\mathbf{1} \le \mathbf{1},\; \mathbf{1}^\top \mathbf{P} \le \mathbf{1}^\top,\; \sum p_{ij} = n_p \right\}.$$1, the paper reports near-perfect overlap on examples including bicycles, motorbikes, guns, and the Eiffel Tower, whereas baseline methods often fail under heavy clutter or occlusion (Lian et al., 26 Mar 2026). The authors interpret this as evidence that the bound better captures global spatial structure.

For 3D synthetic evaluation, the baselines include RPM-HTB, GO-ICP, FRS, GORE, and TEASER++ (Lian et al., 26 Mar 2026). DC-Reg with $$\mathcal{P} = \left\{\mathbf{P} \in \{0,1\}^{N \times M} \;\middle|\; \mathbf{P}\mathbf{1} \le \mathbf{1},\; \mathbf{1}^\top \mathbf{P} \le \mathbf{1}^\top,\; \sum p_{ij} = n_p \right\}.$$2 is reported to have the lowest registration error in deformation and noise tests, and to remain strongest in mixed-outlier settings up to outlier ratio $$\mathcal{P} = \left\{\mathbf{P} \in \{0,1\}^{N \times M} \;\middle|\; \mathbf{P}\mathbf{1} \le \mathbf{1},\; \mathbf{1}^\top \mathbf{P} \le \mathbf{1}^\top,\; \sum p_{ij} = n_p \right\}.$$3 (Lian et al., 26 Mar 2026). The paper is more nuanced for separated outliers and heavy occlusion: for separated outliers, DC-Reg is only moderate and can be outperformed by GO-ICP or GORE; for occlusion plus outliers, it remains stable but is not the best, with TEASER++ and GORE showing stronger resilience to large missing-data regimes (Lian et al., 26 Mar 2026). The authors attribute some of this limitation to the rotation-invariant feature based on norms, which is more vulnerable when clutter is spatially structured.

On the 3DMatch benchmark, using standard D3Feat scan pairs across five RGB-D reconstruction datasets, the paper reports average registration error versus overlap ratio (Lian et al., 26 Mar 2026). The main qualitative conclusion is that DC-Reg maintains low error across datasets, remains particularly stable in low-overlap regimes, significantly outperforms RPM-HTB throughout, and achieves parity with strong robust solvers such as TEASER++ and GORE on real-world data (Lian et al., 26 Mar 2026). No detailed numeric tables are provided in the supplied text, so the most faithful summary remains qualitative.

In the broader registration literature, DC-Reg is best understood as a globally optimal partial-assignment method positioned against deterministic global rigid solvers such as GO-ICP (Chen, 2016), robust correspondence-driven estimators such as TEASER++ (Yang et al., 2020), and outlier-rejection methods such as GORE (Christof et al., 2017). Its distinguishing feature is not merely global search, nor merely correspondence flexibility, but the specific lower-bounding strategy that preserves transformation-assignment coupling within the assignment minimization itself (Lian et al., 26 Mar 2026). This suggests that DC-Reg is particularly relevant where global optimality, poor initialization, partial overlap, and highly ambiguous correspondences are simultaneous requirements.