One-Step Pose Sampling for Cross-Source Point Cloud Registration

Aligning two 3D scans taken by different sensors is a common need in robotics and mapping. Examples include a sparse spinning LiDAR against a dense solid-state LiDAR, or a depth camera against a photogrammetric reconstruction. This problem is called cross-source registration, and it remains brittle. The standard approach finds matching points between the two scans and fits a rigid motion to them. That approach assumes a corner looks the same to both sensors, and across sensors it often does not. On the real-world Cross3DReg benchmark, ICP and the learned matcher Predator reach 0% registration recall (Cross3DReg). A second family of methods avoids explicit matches altogether. It treats the unknown rotation and translation as something to generate, sampled by gradually denoising a random pose while conditioning on both scans (PCRDiffusion, SE(3) Diffusion Model-based Point Cloud Registration). These generators are slow because each sample needs many sequential network passes. Each generator also returns a single answer, even though cross-source pairs with low overlap and asymmetric occlusion often admit several plausible alignments.

We propose to spend the latency savings of one-step generation on breadth instead of speed alone. We would train a pose generator that jumps from noise to a rigid transform in a single network pass, using "average velocity" transport on the space of rigid motions (Riemannian MeanFlow). We would then draw dozens of pose hypotheses in one batched call and keep the one that best explains both scans under a density-normalized overlap check. The result is a learned, correspondence-free replacement for RANSAC's slow sequential hypothesis sampling. Its hypotheses come from a model of which poses are plausible, not from point matches that the domain gap has corrupted.

Core Claim

At a fixed wall-clock budget, K one-step pose samples plus a cheap verification step will register more cross-source pairs than three alternatives. The first is a multi-step flow sampler, which can afford only a few samples in the same time. The second is a deterministic regressor, which gives exactly one answer. The third is a correspondence pipeline, which is limited by the domain gap on exactly those pairs where its matches collapse.

The claim is falsifiable in three places. If a one-step sampler covers fewer alignment modes than a multi-step one, best-of-K accuracy will not improve with K. 3D MeanFlow already warns that one-step sampling of point clouds loses diversity, so this is a real risk and not a formality. If a simple multi-head regressor trained to predict K poses matches our results, the generative machinery is unnecessary. If strong correspondence methods already succeed on nearly every pair, the idea has no room to help.

Relation to Prior Work

Four lines of prior work matter here: generative registration, cross-source registration, one-step transport, and hypothesis selection.

Three papers motivate the generative formulation. PCRDiffusion diffuses a quaternion-plus-translation vector while encoding the scans only once, and reports KITTI rotation RMSE of 0.6281 against 18.1473 for DCP. SE(3) Diffusion Model-based Point Cloud Registration injects noise in the tangent space of the rigid-motion group and truncates to 5 steps. Register Any Point transports points with multi-step flow matching and a rigidity projection, trained on 17 datasets. All three are multi-step, all three return one sample per run, and none is evaluated on real cross-source pairs such as Cross3DReg. We differ in each respect: one network pass per sample, many samples used as competing hypotheses, and training and evaluation on cross-source pairs.

Rectified Point Flow supports the premise that generation handles ambiguity better than regression, since it resolves symmetric assembly poses without averaging them. It still integrates an ODE over several steps and recovers pose afterwards with SVD. We sample poses directly on the rigid-motion group, which also sidesteps the problem of transporting points between scans with very different point counts. Point-wise transport between a scan of $10^5$ points and one of $10^3$ points has no well-defined endpoint.

Two methods supply our one-step mechanism. Mean Flows for One-step Generative Modeling trains a network to predict the average displacement over a time interval rather than the instantaneous velocity, reaching ImageNet FID 3.43 in one step. Riemannian MeanFlow extends this to curved spaces, including the rigid-motion group, and avoids higher-order derivatives. Its evidence that 5-step conventional flows collapse on manifolds (under 10% protein designability, against 82% for RMF) constrains our design: we do not expect naive step truncation of existing registration diffusers to work. That paper studies generation, not conditional registration. The conditioning on two mismatched scans is our contribution.

Two cross-source methods supply the conditioning encoder and the verification idea. Cross-PCR identifies density mismatch as the main failure driver and fuses features across resolutions. It also uses a "loose generation, strict selection" strategy over correspondences. Our design moves that loose-then-strict idea from correspondences to whole poses. SPEAL shows that skeleton-based priors reach 97.3% recall on KITTI CrossSource. It sets the accuracy bar on that benchmark but offers no runtime story.

Generative Point Cloud Registration uses 2D diffusion to hallucinate color for matching. Its authors acknowledge that diffusion cost blocks real-time use. We avoid 2D synthesis entirely.

Research Program

The model has three parts.

The first part is a conditioning encoder that tolerates density differences. We would use the multi-density encoder of Cross-PCR, or GeoTransformer superpoint features, run once per pair. The encoder produces per-scan tokens and a cross-attended pair embedding $c$.

The second part is a small pose head, $u_\theta(g_t, r, t, c)$. Here $g_t$ is a noisy rigid transform, and $r < t$ are two times along the path from noise to the true pose. The head predicts the average geodesic velocity that carries $g_t$ back to time $r$. It is trained with the Riemannian MeanFlow identities on SE(3), using the JVP-free stop-gradient placement. At test time we draw K noise poses and apply the head once each. This gives K candidate transforms in one batched call whose cost is dominated by the shared encoder.

The third part is verification. Each candidate is scored by the fraction of voxel-downsampled source points that fall within a distance threshold of the target. Downsampling both clouds to a common voxel size is what makes the score comparable across a sparse and a dense sensor. The winning candidate can optionally receive one refinement pass through the head.

Training data comes from Cross3DReg, which has 13,231 pairs of mechanical and semi-solid-state LiDAR scans, and from KITTI CrossSource. Training poses are sampled over wide initial misalignments so that the learned distribution over poses genuinely spreads when overlap is low.

We would also examine a by-product: the spread of the K samples as a failure detector. Cross-source registration usually fails silently. A method that can say "these scans admit no confident alignment" is useful on its own. We would measure this by the area under the ROC curve of sample spread for predicting failure.

Baselines

The naive reference points are ICP and Predator on Cross3DReg, both reported at 0% recall (Cross3DReg), and HRegNet at 69.3% recall on KITTI CrossSource (SPEAL). We would rerun these under our protocol. They establish the regime in which correspondence-free hypotheses can matter.

The strongest existing methods differ by benchmark. On Cross3DReg it is the image-guided method of Cross3DReg, which reports 87.1% registration recall, described as 5.4% above GeoTransformer. On KITTI CrossSource it is SPEAL, at 97.3% recall against 96.8% for GeoTransformer. Cross-PCR is the strongest geometry-only method, but on 3DCSR Kinect-LiDAR the file gives only a relative gain (+57.6 percentage points), so no absolute recall value is available on Lacuna for that benchmark. No reported runtime was found on Lacuna for Cross-PCR, SPEAL, GeoTransformer, or RAP, so we would measure all of them ourselves on one GPU.

Because the image-guided method uses RGB, we would report it separately and compare geometry-only methods against GeoTransformer and Cross-PCR retrained on the same splits. Beating the strongest baseline means one of two things. The first is matching its recall at lower end-to-end latency. The second is exceeding its recall at equal latency, with the gain concentrated on the subset of pairs where the correspondence baseline fails.

Four ablations isolate the key ingredient. All share our encoder, data, training epochs, and a fixed budget of hyperparameter trials. The first is a direct regression head, which gives one pose in one pass. The second is a multi-head winner-take-all regressor that outputs K poses. The third is a multi-step SE(3) flow-matching head at 10 to 50 steps, which serves as the accuracy ceiling for sampling. The fourth is our one-step head with K = 1, which removes selection. The claim survives only if K-sample one-step selection beats the first, second, and fourth ablations at matched latency, and approaches the third at a fraction of its latency.

Why not just…?

Why not just regress the pose in one pass? Regression is as fast as our method and far simpler. The file's evidence against it is weak because it is confounded: DCP's KITTI rotation RMSE of 18.1473 against PCRDiffusion's 0.6281 mixes the effect of the head with that of the backbone. Our first ablation removes the confound by using an identical encoder. We expect regression to fail mainly on low-overlap pairs whose correct pose is ambiguous, because averaging two plausible poses produces an invalid one, which is the failure mode that Rectified Point Flow documents for symmetric parts. If regression matches our method on those pairs, we will drop the generative framing.

Why not just predict K poses with a multi-head winner-take-all regressor? This is the most serious threat. It delivers multiple hypotheses without any flow mathematics. Its known weakness is that the heads can collapse or leave some heads unused, and the number of modes it can represent is fixed at training time. A sampler, by contrast, can draw more hypotheses at test time. The second ablation tests this directly. The deciding plot is best-of-K error as K grows beyond the number of heads. If the multi-head regressor matches our method there, the honest conclusion is that multi-hypothesis prediction, not average-velocity transport, is the useful ingredient.

Why not just use GeoTransformer or Cross-PCR with RANSAC? These methods are strong. On KITTI CrossSource there is almost no headroom, since GeoTransformer already reaches 96.8%. We therefore do not expect to win on average recall there. The argument is about the pairs where correspondences collapse and RANSAC's sample count explodes at low inlier ratios. We would stratify every benchmark by overlap and by baseline inlier ratio, and report results per stratum. We would also test a hybrid that adds our hypotheses to RANSAC's pool.

Why not just distill RAP or truncate an existing registration diffuser to one step? Truncation is the cheapest option, but Riemannian MeanFlow's 5-step collapse result predicts it will degrade badly on SE(3). We would verify this by running PCRDiffusion-style and flow-matching heads at 1 and 2 steps. Distillation needs a trained multi-step teacher plus several additional training stages. Our training cost would be matched against it and reported.

Proposed Experiments

The main experiment is an accuracy-versus-latency Pareto curve on Cross3DReg and KITTI CrossSource. Accuracy is measured by registration recall, relative rotation error, and relative translation error. Latency is end-to-end wall-clock time including the encoder, on one GPU, with K swept from 1 to 128. The multi-step flow head and correspondence baselines appear as points on the same axes. We would add GermanyForest3D binned by overlap, and 3DCSR Kinect-LiDAR as a smaller transfer test with no retraining.

The first decisive check is mode coverage. Best-of-K rotation error should fall with K for the one-step sampler about as steeply as for the 20-step sampler. If it plateaus early, the diversity loss noted in 3D MeanFlow has defeated the design.

The second decisive check is recall on the stratum of Cross3DReg pairs where retrained GeoTransformer fails. This is where a correspondence-free generator has to earn its place.

The third decisive check is whether sample spread predicts failure, measured as the area under the ROC curve over all test pairs.

An extension to scale-ambiguous Kinect-SFM pairs, which requires the similarity group with a scale parameter, is deferred until the SE(3) results hold.