Deep Global Registration Learning

Hand descriptors break on sparse noisy scans. Deep global registration learns matching features plus a safe solver for wide-baseline snaps.

Learned descriptors propose matches that a robust solver turns into one pose

Why Does This Exist?

Classical point cloud registration with FPFH plus RANSAC collapses under low overlap and changing density. Deep global registration exists to learn both halves: fully convolutional descriptors that survive rotation and sparsity, plus a differentiable-safe solver that down-weights the remaining outliers. It handles the wide-baseline pairs where ICP cannot start.

Local polish stays ICP; the problem map is registration.

Think of It Like This

Bilingual guides plus a strict judge

Hire local guides who recognize landmarks in any weather, then let a strict judge discard flaky stories and rule on the shared map. Better guides plus a sterner judge beat loud witnesses.

Descriptors are guides and the robust solver is the judge. The analogy stops at training: guides learn from labeled scans while humans generalize without datasets.

How It Actually Works

A sparse-convolution backbone outputs per-point descriptors such as FCGF trained with contrastive loss on matching versus non-matching pairs. Matching proposes thousands of correspondences; a learned outlier mask or deep robust estimator weights them, then weighted SVD or RANSAC solves $R$ and $t$ . Metrics report recall under $10$ cm and $5$ degrees on hard indoor pairs with $30$ percent overlap.

Worked numbers

On a hard split, FPFH plus RANSAC recalls $70$ percent of pairs under $10$ cm. Learned FCGF plus the robust head recalls $92$ percent. Inlier ratio rises from $8$ percent to $24$ percent, cutting RANSAC rounds for $99$ percent confidence from $8993$ to $331$ .

Code

1def ransac_rounds(inlier: float, need: int = 3, conf: float = 0.99) -> int:2    import math3    return math.ceil(math.log(1 - conf) / math.log(1 - inlier ** need))4
5
6print(ransac_rounds(0.08))  # -> 89937print(ransac_rounds(0.24))  # -> 331

Watch Out For

Deploying across sensors without retraining

Descriptors tuned on RGB-D rooms degrade on spinning lidar streets because density and noise differ. The symptom is recall crashing from $90$ to $50$ percent on the new sensor. Finetune on the target sensor or keep a classical fallback for cross-device pairs.

The Quick Version

Deep global registration learns descriptors plus robust solving together.
Contrastive training makes matches survive rotation and sparsity.
Learned masks raise inlier ratios from $8$ to $24$ percent on hard pairs.
Recall under $10$ cm rises near $92$ percent versus $70$ classical.
RANSAC rounds fall from $8993$ to $331$ from better inliers.
Sensor change without retraining is the main deployment trap.