One-Step Pose Posteriors for Camera-to-Satellite Localization

When GPS fails in a dense city, a car or robot can still find itself by matching what its camera sees against a satellite photo of the area. Today's best matchers return a single answer, and they fail in a specific, repeatable way. On long, symmetric streets they confidently report a heading that is exactly backwards, because from the camera's point of view the street looks nearly the same in both directions. This proposal treats localization as drawing many samples from a probability distribution over the vehicle's position and heading, so that "here, facing north" and "here, facing south" can both survive until evidence separates them. The obstacle is speed. Generators that turn noise into samples usually need dozens of sequential steps, and here each step must re-read the satellite image at hundreds of candidate poses.

The plan is to train a flow model, meaning a generator that learns a velocity field carrying random samples to data, directly on the space of planar poses (position plus heading). We use the recent "average velocity" formulation, which jumps from noise to a sample in one to four network calls instead of fifty. In the resulting localizer, each call checks every candidate pose against the satellite image. The output is a set of a few hundred pose hypotheses per frame, produced faster than current single-answer methods. That set can be fused over time with odometry.

Core Claim

Average-velocity transport on the group of planar rigid motions, written SE(2), can turn a wide, uninformative pose prior into a calibrated, multi-modal posterior over $(x, y, \theta)$ in one to four function evaluations, where $x, y$ is position and $\theta$ is heading. This should do two things a deterministic refiner cannot. It should reduce 180-degree heading flips, measured as a lower top-1 flip rate and a high rate of having the true pose among the top two modes. It should also localize from priors much wider than the ±30 m capture range of current iterative refiners, while keeping latency below theirs. The new part is not using a generative model for cross-view work, which has been done for retrieval. The new part is making metric pose itself the generated variable, with every flow step conditioned on satellite features read out at the current pose hypothesis.

Positioning Against Prior Work

Two papers motivate the gap. OpenCVL builds a 617k-pair, 41-city benchmark and reports that strong models suffer "180-degree orientation flips" in repetitive corridors. PIDLoc refines 3-DoF pose with control-inspired branches. It shows that a purely proportional update fails beyond 10 m of initial error and that a grid-search branch is needed to reach ±30 m, at 370 ms. Both results indicate that a single point estimate is a poor representation when the posterior has several peaks.

Three lines of deterministic pose recovery constrain the design, because they set the accuracy bar. FG² lifts DINOv2 features into height-pooled pillars and solves pose by differentiable Procrustes alignment, but it is sensitive to the initial heading. Learning Dense Flow Field predicts dense pixel displacement with a RAFT-style recurrent unit and solves pose by weighted least squares. Its "flow" is optical flow in image space, not probability transport. VIGOR supplies the non-centered tile setting and a retrieve-then-regress pipeline. We keep the feature machinery of these methods and replace their single-answer output stage.

The existing generative cross-view work does not generate poses. DiffusionUavLoc conditions a ControlNet on pseudo-orthophotos and avoids sampling by reading descriptors from the autoencoder latent, so it performs retrieval only. MCGF restores weather-degraded images with a multi-step diffusion model before retrieval and itself reports latency as the barrier. Neither outputs metric coordinates.

The method comes from the fast-sampling literature. Mean Flows trains on average rather than instantaneous velocity and reaches 1-step ImageNet FID 3.43. Riemannian MeanFlow extends this to curved spaces including SE(3), with a stop-gradient objective that needs no Jacobian-vector products. It reports 82% protein designability at 5 steps, where multi-step frame baselines fall below 10%. It has never been used with sensor-conditioned pose.

Both MeanFlow follow-ups also contain a warning that directly threatens our claim. 3D MeanFlow and Riemannian MeanFlow report that one-step samplers lose diversity and favor dominant modes, which here could mean erasing the minority heading mode. We treat this as the central risk and test it directly.

Research Program

The localizer has three parts.

The first part is a shared encoder, initialized from DINOv2 as in FG². It lifts the ground image to a bird's-eye-view feature map and encodes the satellite tile once per frame.

The second part is a velocity network $u_\phi(g, r, t)$. Here $g \in SE(2)$ is the current pose hypothesis, $r < t$ are two times in $[0,1]$, and $\phi$ are the network weights. At each call, the network warps the ground feature map into the satellite frame at $g$ and computes a small local correlation patch around it. Together with a pooled global embedding, this patch feeds a small network that outputs a velocity in the Lie algebra of SE(2), the three-dimensional space of planar translational and rotational velocities. A hypothesis is moved forward with the group exponential:

$$g_t = g_r \cdot \exp\big((t - r)\, u_\phi(g_r, r, t)\big),$$

where $\exp$ maps a planar velocity to a rigid motion. The one-step sampler sets $r=0$, $t=1$. Few-step samplers chain such jumps, using the semigroup property from Riemannian MeanFlow.

The third part is the source distribution. It is uniform over translation within the satellite tile and uniform over heading, so the method does not depend on an initial guess. At test time we push $K=256$ samples through the network in parallel. Training uses the JVP-free conditional Riemannian MeanFlow objective, with geodesic interpolation on SE(2) between source and ground-truth pose.

To guard against mode loss, we add two cheap correctives and ablate both. The first is two to four NFE refinement, where NFE means number of function evaluations, i.e. network calls. The second is a Metropolis-Hastings correction step adapted from AC-Sampler. It uses an auxiliary pose-likelihood head trained on the same correlation features.

For image sequences, the per-frame sample sets are propagated through visual odometry and reweighted by the next frame's kernel density estimate, a standard particle-filter fusion. This lets a flip that is ambiguous in one frame be resolved by later frames.

Baselines

The do-nothing reference is to report the prior as the answer. This means the tile center with random heading on VIGOR and OpenCVL, and the perturbed initial pose on KITTI. It gives the error floor that any learned method must beat by a wide margin.

The strongest existing metric method on KITTI is PIDLoc, which reports a cross-area mean position error of 4.96 m and 370 ms runtime, and stays stable up to ±30 m initial error. We will rerun it from the released configuration on our splits and hardware. We will evaluate it at its native ±30 m prior and at wider priors of ±60 m and full tile, where its grid branch must grow. Beating PIDLoc means matching or improving 4.96 m at ±30 m in under 370 ms on the same GPU, and clearly outperforming it at wider priors.

FG² (7.45 m KITTI cross-area; 2.41 m VIGOR cross-area) and the dense-flow method (0.47 m KITTI same-area median) are run with their own training recipes on identical data, with the same encoder budget where swappable. On OpenCVL the reference is Loc² (7.90 m mean error in the wild); no reported flip rate was found on Lacuna for any method, so we will measure it for all baselines ourselves.

Three ablations use our exact encoder, data, and tuning budget. The first replaces the flow with a mixture-density head that outputs eight weighted SE(2) Gaussians. The second is the same velocity network trained with standard flow matching and sampled at 50 to 100 steps. The third removes the pose-conditioned correlation patch, leaving only global features, to test whether re-reading the satellite map at each hypothesis is what carries the accuracy.

Why not just…?

Why not just run a deterministic refiner twice, at θ and θ+π? This handles exactly the flip ambiguity, and it is cheap. It does not handle translational aliasing along repeated blocks or priors wider than the capture range, both of which need many hypotheses. We include "PIDLoc ×2 headings" as a baseline. If it matches our flip rate and accuracy at ±30 m, the proposal's claim narrows to wide-prior global localization, and we will report that honestly.

Why not just predict a mixture of Gaussians in one pass? This is the most serious threat. It is multi-modal, single-pass, and simple. Its weakness is that a fixed set of components must be predicted from global features without checking each hypothesis against the map. The ablation with identical encoder and calibration metrics settles the question. If the mixture head matches our method, the flow model is not needed.

Why not just use ordinary flow matching, since pose is only three-dimensional? Each step is cheaper than the image encoder, so many steps might be affordable. However, each step evaluates a correlation patch for all 256 hypotheses, so per-step cost scales with the number of hypotheses. Riemannian MeanFlow also reports that multi-step frame models collapse at low step counts. We will plot accuracy and calibration against wall-clock time for 1, 2, 4, 10, 50, and 100 NFE. If 10-step flow matching fits the latency budget with better quality, we adopt it and drop the one-step claim.

Why not just denoise the images with a diffusion model, as in MCGF? That approach addresses weather, not ambiguity, and still yields retrieval rather than pose. It is complementary to this proposal, and we leave it out of scope.

Proposed Experiments

We train and evaluate on KITTI (same-area and cross-area), VIGOR (same-area and cross-area), and OpenCVL (cross-area, snowy, and in-the-wild splits). Ford Multi-AV Seasonal is held out as an untouched transfer test. The metrics are mean and median position error, median heading error, flip rate (top mode more than 150° off), top-2-mode recall, coverage of the 50%, 90%, and 95% highest-density regions as a calibration measure, and end-to-end latency at batch one.

There are three decisive checks. The first is mode preservation. On a controlled set of OpenCVL corridor frames, which we augment with their 180°-rotated satellite counterparts so that the true posterior is known to be bimodal, we measure the mass assigned to the minority mode as a function of NFE. If one-step sampling drops it below 20% while four steps recover it, the claim survives only at four steps. The second is capture range. We plot error against prior width from ±10 m to the full tile, comparing against PIDLoc. The third is temporal fusion on KITTI sequences. We test whether the frame-level flip rate falls toward zero within a few frames, compared against single-answer methods fused with the same odometry.