Minimal Solvers for Mini-Loop Closures in 3D Multi-Scan Alignment

1. Introduction

• Higher-order RGB-D pose estimation had received no attention, despite progress on RGB-camera problems and the practical correspondence counts for n≤5.The introduction points to algebraic minimal solvers for depth sensors as a needed addition.Source: §1 · p. 1
• Seven point matches jointly register four scans and recover 18 transformation variables, while pairwise methods can accumulate drift.The four-scan example uses a 4-cycle minimal solver.Source: §1 · p. 1, Figure 1 · p. 1

1.1. Related Work

• Related work spans ICP and other 3D alignment methods, geometry-based SLAM, registration with semantic segmentation, and deep-learning approaches.The introduction also notes surveys of 3D SLAM evaluation and rigid and non-rigid point-cloud registration.Source: §1.1 · p. 2

1.2. Notation and Problem Definition

• An n-cycle is a sequence of scans with loop closure between the first and last point clouds; the goal is to find transformations between scan coordinate systems.The notation section represents these transformations in SE(3).Source: §1.2 · p. 2

1. Introduction

• The paper proposes minimal solvers for 3-, 4-, and 5-cycles in general six-degree-of-freedom and planar-motion settings.The authors describe these cases as novel to the best of their knowledge.Source: §1.2 · p. 2

2. Minimal Solvers

• Predefined transformations align scan frames around virtual axes through matched points, reducing the registration variables to rotations around those axes.After solving, inverse predefined transformations recover poses in the original coordinate frames.Source: §2 · p. 3

2.1. Setting the Stage for Minimal Solvers

• Two point correspondences between each consecutive scan pair define the virtual axes, leaving one rotation angle per pair to estimate.For three scans, Figure 2 depicts the remaining degrees of freedom after these transformations.Source: Figure 2 · p. 3, §2.1 · p. 4

2.2. Pairwise Registration

• For pairwise registration, a third point correspondence yields a closed-form solution, with up to two candidates when noisy data violate the trigonometric constraint.The candidates preserve a rotation matrix and can be checked by back-substitution.Source: §2.2 · p. 4

2.3. 3-Cycle Registration

• For a 3-cycle, two residual angles are solved separately using one closing correspondence, producing up to four valid solutions.Each closure point correspondence supplies three constraints for the two unknown angles.Source: §2.3 · pp. 4–5

2.4. 4-Cycle Registration

• For relative-pose recovery, three equations in three unknowns mean a single point correspondence can generally yield a solution; the automatic Gröbner-basis generator returns up to 16 solutions.The solver combines three constraints with three trigonometric constraints.Source: §2.4 · p. 5

2.5. 5-Cycle Registration

• The general n-cycle formulation uses two correspondences to close the loop and yields 32 solutions for n = 5; the authors report it feasible only through n = 5.For n = 6, it may yield up to 288 solutions, with no easy way to build the solver.Source: §2.5 · p. 5

3. Planar Motion Case

• Planar motion has 3 degrees of freedom: 2 translations and 1 rotation.Source: §3 · p. 5
• Choosing the z-axis as the rotation axis makes the second correspondence in the general construction unnecessary, leaving one 3D correspondence per scan pair.Source: §3 · pp. 5–6
• The planar-motion table summarizes the minimum correspondence counts for n-cycles; because pairwise registration requires two points, the paper considers cycles of at most four scans.Source: Table 2 · p. 6, §3 · p. 6

4. Motion Averaging

• Pose-graph edges connect camera pairs with scene overlap, identified using SURF matches on RGB images and a minimum feature-correspondence threshold.Source: §4 · p. 6
• A depth-first search extracts edge-disjoint cycles, prioritizing 5-cycles, then 4-cycles, then 3-cycles; remaining edges use pairwise registration.Source: §4 · p. 6
• The solver timing table reports milliseconds for a Matlab implementation; the caption notes that C++ would speed up computation.Source: Table 3 · p. 6
• Redundant relative poses are averaged: an L1 rotation solution initializes IRLS, after which translations are computed by linear least squares.Source: §4 · p. 6

5. Experimental Results

• The experiments include small n-cycle graphs and larger datasets decomposed into edge-disjoint cycles, with error evaluated against ground truth.Source: §5 · p. 6

5.1. Synthetic Data

• The synthetic study tests five scans with pairwise, 3-cycle, 4-,2-cycle, and 5-cycle approaches, using 1000 RANSAC iterations and a 50-unit inlier threshold.Source: Figure 4 · p. 7, §5.1 · p. 7
• In synthetic trials, n-cycle solvers reduce overall rotation and translation error; the 5-cycle has the lowest errors but also the lowest inlier count.Source: §5.1 · p. 7, Figure 4 · p. 7

5.2. Real Experiments

• Real-data evaluation uses three TUM sequences with ground-truth camera positions, SURF matches on RGB images, and corresponding depth points.Source: §5.2 · p. 7
• On TUM five-scan sequences, n-cycle methods generally outperform pairwise registration, and the 5-cycle yields better rotation errors.Table 4 also reports how often n-cycles outperform Pairwise.Source: §5.2 · p. 7, Table 4 · p. 8
• For the larger TUM pose graphs, rotation averaging produces camera rotations and least squares estimates translations; the final registrations show 100 scans from each sequence.Source: §5.2 · p. 8, Figure 5 · p. 8

• For four cameras, the n-cycle method uses 7 point correspondences, versus a minimum of 9 for pairwise methods.Source: §6 · p. 8
• Table 5 reports solver edge coverage and average rotation and translation errors for the registrations shown in Fig. 5.Source: Table 5 · p. 8, §6 · p. 8
• The authors characterize their n-cycle solvers as exactly minimal, not near-minimal.Source: §6 · p. 8
• The solvers provide alternate relative poses that may benefit pose-graph refinement or motion averaging, but solving n-cycles with n > 5 was not practically feasible.Source: §6 · p. 8