Aligning two 3D scans means finding the rotation and translation that place one on top of the other. Neural registration networks now do this well even when the scans barely overlap, and much of their accuracy comes from iterative refinement: the network predicts a pose, re-examines the scans under that pose, and corrects it, over several passes. Every pair pays for every pass, whether it was already aligned after the first pass or still badly wrong after the last. This note proposes using a cheap, non-learned geometric test, borrowed from classical outlier rejection, to decide after each pass whether the current pose is already correct. The test asks whether many point matches agree with the pose and also agree with each other under distance preservation. The refinement loop then spends extra passes only on pairs that need them.
In plain terms, iterative registration networks run a fixed number of expensive passes. Geometric outlier pruning can check a pose in milliseconds without any learning. We put the second inside the first as a stopping rule and as a source of a better starting pose, and we measure whether average latency falls while accuracy holds. The idea stands or falls on one question that nobody in the file has measured: what fraction of refinement passes are wasted on pairs that were already solved.
Core Claim
Rigid motion preserves distances. Two correct matches $(p_a,q_a)$ and $(p_b,q_b)$ must satisfy $\big|\|p_a-p_b\|-\|q_a-q_b\|\big|\le 2\epsilon$, where $\epsilon$ is the sensor noise bound. You can check this without knowing the pose. Graph methods build a "compatibility graph" whose nodes are candidate matches and whose edges connect pairs that pass this test. True matches then form a densely connected cluster: a maximum $k$-core in ROBIN, maximal cliques in MAC, or a $k$-truss in PointTruss. A $k$-truss is a subgraph in which every edge lies in at least $k-2$ triangles.
Our claim is that this structure can serve as a pose certificate inside a refinement loop. At step $t$ the network outputs pose $T_t$ and correspondences $\mathcal{C}_t$. We take the matches whose residual under $T_t$ is below $\tau$ and keep only those that survive $k$-truss filtering on the compatibility graph restricted to them. Call this certified set $\mathcal{S}_t$. We stop when $|\mathcal{S}_t|$ exceeds a threshold and the closed-form pose fitted to $\mathcal{S}_t$ agrees with $T_t$ to within a small rotation and translation tolerance. If we do not stop, $\mathcal{S}_t$ supplies two things to the next pass: the overlap anchors, in place of a mask derived only from the previous pose, and a warm-start pose.
Restricting the graph to matches already consistent with $T_t$ matters for speed. It shrinks the graph from thousands of nodes to roughly the inlier count, which sidesteps the quadratic graph-construction cost and the density blow-up that took MAC to 331 s on ETH "trees" in the TCF comparison.
Relation to Prior Work
Our method builds on the refinement networks. AMR replaces PEAL's single network, applied for five steps, with step-specialized networks. Its two-step version beats five-step PEAL, reaching 80.4% registration recall on 3DLoMatch. However, AMR still runs a fixed number of steps on every pair, and it selects overlap anchors from the previous pose with no hard geometric check. GeoTransformer builds distance and angle invariants into its features and solves pose in one closed-form pass, reaching 75.0% on 3DLoMatch. The gap between 75.0% and AMR's 80.4% is the evidence that refinement helps.
Our method takes its machinery from the pruning work. We reuse PointTruss's $k$-truss filtering and its spatial-distribution score, which penalizes clusters of agreeing matches that sit in one small region. We also reuse TCF's cascade of one-point and two-point checks as a cheaper alternative certificate. The pruning literature also limits what we can claim. MAC shows that in low overlap the largest consistent cluster can be spurious, and the file notes that repetitive or symmetric scenes produce large consistent clusters of wrong matches. A geometric certificate can therefore approve a wrong pose with confidence. This is the main risk we measure.
Finally, the file lists the lack of principled early stopping as an open problem in iterative refinement generally. DCFold attacks the same latency problem in protein folding the opposite way, by distilling iterations into one pass. We keep the iterations and spend them selectively per input, which avoids retraining.
Research Program
We will instrument AMR and a PEAL-style static loop, using released weights where available and retraining otherwise. For every test pair and every step, we will log the pose, the correspondences, and whether the pose meets the benchmark's success criterion.
The first analysis comes before any method work: an oracle stopping curve that stops each pair at the first step whose pose is already correct. This gives an upper bound on the achievable reduction in average passes. If that bound is small, the project ends early with a useful negative result.
We then implement three certificates of increasing strength:
- a residual inlier count under $T_t$;
- the same count after TCF-style one-point and two-point cascade filtering;
- the full $k$-truss certificate with the spatial-distribution score.
We will calibrate thresholds on the 3DMatch training scenes only and report a curve of false-stop rate (stopped on a wrong pose) against passes saved.
Separately, we test feeding $\mathcal{S}_t$ forward as anchors and as a warm start. Some pairs may then succeed in fewer steps rather than merely stop earlier.
Baselines
The naive reference is AMR or PEAL run for its fixed number of steps. This gives full accuracy at full cost and defines the latency we must reduce. For accuracy we compare against AMR's reported 80.4% registration recall on 3DLoMatch from AMR. We found no reported per-pair latency for AMR on Lacuna, so we will measure it ourselves on the same GPU.
The strongest non-iterative competitor is GeoTransformer correspondences followed by a graph solver. MAC reports 78.9% on 3DLoMatch and 95.7% on 3DMatch with GeoTransformer descriptors in MAC, and PointTruss reports 99.64% recall at 0.21 s on KITTI. These pipelines run once, so if they come within a point of AMR at lower latency, iteration itself is questionable.
The strongest early-stopping competitor is a pose-change rule: stop when $T_t$ moves less than a set rotation and translation from $T_{t-1}$. A second is a rule based on the network's own predicted overlap or matching confidence.
To isolate the key ingredient, mutual geometric consistency, we include an ablation that uses only the residual count, with no compatibility-graph filtering.
All methods use the same correspondences and the same hardware. Every stopping threshold gets the same calibration budget on 3DMatch training scenes. Beating the field means matching fixed-step AMR's recall within 0.3 points on 3DLoMatch while cutting mean wall-clock time per pair by a meaningful fraction, and doing so at a lower false-stop rate than the pose-change and confidence rules at equal passes saved.
Why not just…?
Why not just stop when the pose stops changing? This is the most natural rule and costs nothing. Its failure mode is a confident stall: a low-overlap pair locked in a wrong local minimum also stops moving. DiffusionReg's motivation, iterating to escape local minima under occlusion, suggests such stalls occur. Whether they are common enough to matter is unknown, so the false-stop curves will decide it directly. If the pose-change rule matches our certificate, we will report that and drop the geometry.
Why not just trust the network's own confidence? A learned confidence costs nothing extra. However, it is trained on the same features that produced the error, so it tends to be overconfident exactly on repetitive or ambiguous structure. The certificate is independent evidence, because it depends only on raw coordinates. The direct comparison is in the Baselines section. We will also break results down by overlap bin, since the gap should be widest at 10–30% overlap.
Why not just run MAC or PointTruss once on GeoTransformer matches and skip iteration? This is a real threat. MAC's 78.9% is within 1.5 points of AMR's 80.4% on 3DLoMatch, possibly at lower cost. Our method uses the same graph test, but on only the matches an iterative pose has already made plausible. That should keep the robustness gains of refinement while avoiding the dense-graph cost, which reached 331 s in the TCF comparison. If the one-shot pipeline matches adaptive AMR on both recall and latency, the conclusion is that iteration is unnecessary on these benchmarks, and we will state it.
Why not just use AMR with fewer fixed steps? AMR's two-step version already does this. The question is whether per-pair variation in difficulty is large. The oracle curve answers it: if most pairs need exactly the same number of steps, a fixed small budget is the right answer and adaptivity adds nothing.
Decisive Experiments
We evaluate on 3DMatch, 3DLoMatch, and KITTI using registration recall under each benchmark's standard thresholds. We also report mean and 90th-percentile per-pair latency, mean passes, and false-stop rate. ETH "trees" serves as a stress test of certificate cost under dense, low-inlier conditions. Synthetic Stanford Bunny pairs with injected symmetric outliers serve as a stress test of certificate correctness.
Three results would be decisive. First, the oracle curve on 3DLoMatch must show that a substantial share of pairs are already solved after step one. Second, the $k$-truss certificate must beat both the pose-change and confidence rules on the false-stop-versus-savings curve, especially in the lowest overlap bin. Third, adaptive AMR must keep recall within 0.3 points of 80.4% while beating MAC+GeoTransformer on latency at equal or better recall.
The symmetric-scene test is where we expect the certificate to fail. We will report its false-accept rate there openly, along with whether the spatial-distribution score reduces it.
References
- ROBIN: a Graph-Theoretic Approach to Reject Outliers in Robust Estimation using Invariants
- 3D Registration with Maximal Cliques
- PointTruss: K-Truss for Point Cloud Registration
- RANSAC Back to SOTA: A Two-Stage Consensus Filtering for Real-Time 3D Registration
- Adaptive Multi-step Refinement Network for Robust Point Cloud Registration
- Geometric Transformer for Fast and Robust Point Cloud Registration
- DCFold: Efficient Protein Structure Generation with Single Forward Pass