Noisy Correspondence in Cross-View Geo-Localization

• The paper introduces NC-CVGL by formulating noisy correspondences from GPS drift as a partial overlap challenge and employing uncertainty-guided partition and augmentation.
• NC-CVGL is defined as a robustness problem in cross-view geo-localization that addresses spatial misalignment, dynamic scene variations, and view-dependent clutter.
• The PAUL framework leverages dual peer networks with co-divide, co-augment, and co-training stages to extract reliable supervision from semi-positive pairs, enhancing localization metrics under noise.

Noisy Correspondence on Cross-View Geo-Localization (NC-CVGL) is a cross-view geo-localization problem setting in which nominally matched image pairs are not perfectly aligned and may overlap only partially. The term was formally introduced in "PAUL: Uncertainty-Guided Partition and Augmentation for Robust Cross-View Geo-Localization under Noisy Correspondence" (Li et al., 27 Aug 2025), which attributes the phenomenon to practical GPS drift and treats many training pairs as semi-positive rather than fully corresponding. In the broader cross-view geo-localization literature, closely related noise mechanisms include inaccurate geo-poses relative to aerial imagery (Fervers et al., 2022), dynamic or off-the-ground objects and seasonal variation (Wang et al., 2023), irrelevant edge activations in cross-view object geo-localization (Zhu, 31 Oct 2025), and view-specific noisy signals addressed by an information-bottleneck formulation (Zhang et al., 8 May 2026). Taken together, these works define NC-CVGL as part of a larger effort to make cross-view matching robust to partial overlap, label inaccuracy, scene dynamics, and view-dependent clutter.

1. Definition and conceptual scope

In the formal PAUL definition, NC-CVGL arises when imperfect GPS induces systematic alignment shifts between UAV images and satellite images, so that a matched pair may only partially overlap (Li et al., 27 Aug 2025). This differs from the idealized assumption, common in earlier embedding-based CVGL pipelines, that each positive pair is perfectly aligned during training. The resulting supervision noise is structurally important: the pair is not necessarily a false match, but neither is it a clean positive.

Related work shows that the phrase “noisy correspondence” covers multiple technical failure modes rather than a single artifact. In metric CVGL, inaccurate ground-truth geo-poses relative to aerial tiles degrade supervision and motivate pseudo-label refinement and uncertainty-aware pose estimation (Fervers et al., 2022). In sparse feature-based localization, correspondences are corrupted by dynamic objects, off-the-ground objects, and seasonal changes, prompting explicit purification mechanisms (Wang et al., 2023). In cross-view object geo-localization, attention modules are designed to suppress “edge noise”—activations that are visually salient but geometrically irrelevant (Zhu, 31 Oct 2025). In object-centric retrieval, the problem is recast as separating view-invariant structural information from view-specific noisy signals (Zhang et al., 8 May 2026).

A useful distinction follows from these formulations. NC-CVGL in the strict PAUL sense is a pair-level spatial misalignment problem caused by partial overlap, whereas several adjacent methods handle correspondence-level or representation-level noise. This suggests that NC-CVGL is best understood as a family of robustness problems inside CVGL rather than a single annotation error model.

2. Formal problem formulation

PAUL defines the query and reference sets as

$$\mathcal{Q}=\{q_i\}_{i=1}^N,\qquad \mathcal{R}=\{r_j\}_{j=1}^M,$$

where $\mathcal{Q}$ contains UAV images and $\mathcal{R}$ contains satellite images, each with geographic tags (Li et al., 27 Aug 2025). Overlap is quantified by

$$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$

Two subsets of positive training pairs are then defined by thresholding IoU:

$P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$

$N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$

with $\tau_m=0.39$ for “well-aligned” pairs and $\tau_s=0.14$, below which a pair is treated as negative. The key point is that $N$ contains semi-positive pairs: the images are geographically related, but only partially corresponding.

Observed labels are binary, $y_{ij}\in\{0,1\}$, but PAUL introduces an unobserved latent indicator $\mathcal{Q}$0 specifying whether a nominal match is actually noisy. The learning objective is to obtain an embedding

$\mathcal{Q}$1

that pulls true positives together and pushes negatives apart while remaining robust to the unobserved $\mathcal{Q}$2:

$\mathcal{Q}$3

This formalization is narrower than generic noisy-label learning. The noise is not arbitrary class corruption; it is spatially structured and tied to the geometry of partial overlap. That distinction matters because robust learning can exploit the still-valid subregions of a semi-positive pair rather than discarding the pair entirely.

3. Methodological landscape

Several methodological lines converge on NC-CVGL and adjacent noise-robust CVGL problems.

Work	Noise source addressed	Main mechanism
(Fervers et al., 2022)	Inaccurate ground truth poses	Uncertainty-aware pose distribution and pseudo-label graph optimization
(Wang et al., 2023)	Dynamic objects, off-the-ground objects, seasonal changes	View-consistent confidence, on-ground confidence, spatial embedding
(Li et al., 27 Aug 2025)	Partial overlap from GPS drift	GMM partition, EDL-guided masking, peer co-training
(Zhu, 31 Oct 2025)	Irrelevant edge noise	Iterative cross-attention and multi-scale spatial attention
(Zhang et al., 8 May 2026)	View-specific noisy signals	Information bottleneck with object-centric structural alignment

The uncertainty-aware metric CVGL model of Oliveira et al. predicts a posterior over discrete pose hypotheses $\mathcal{Q}$4 using a vision-only PV$\mathcal{Q}$5BEV$\mathcal{Q}$6matching pipeline (Fervers et al., 2022). Because public geo-poses are often inaccurate with respect to aerial images, the method introduces a pseudo-label pipeline: a LiDAR-based warm-up model predicts poses and covariances; consecutive-frame odometry and aerial-pose unaries form a pose graph; global least-squares optimization via g²o refines the poses; and the worst $\mathcal{Q}$7 of frames are removed according to generalized variance. This is not presented as NC-CVGL, but it directly addresses noisy geometric supervision.

"View Consistent Purification for Accurate Cross-View Localization" (Wang et al., 2023) treats noise as feature-level contamination from scene dynamics and geometry violations. Its Spatially Aware Feature and Confidence Extractor predicts multi-scale feature maps together with two confidence maps: a view-consistent confidence $\mathcal{Q}$8 and an on-ground confidence $\mathcal{Q}$9. Their fusion suppresses points that either do not appear in both views or violate the ground-plane assumption. The method then performs sparse keypoint matching and Levenberg–Marquardt pose refinement.

AttenGeo extends the noise discussion into cross-view object geo-localization (Zhu, 31 Oct 2025). Its Cross-view and Cross-attention Module alternates bidirectional CAB updates for $\mathcal{R}$0 iterations, with empirical evidence that $\mathcal{R}$1 maximizes [email protected] on CVOGL and G2D. A Multi-head Spatial Attention Module uses kernel sizes $\mathcal{R}$2, $\mathcal{R}$3, and $\mathcal{R}$4 to refine fused feature maps and suppress residual false positives around edges.

InfoGeo reframes robustness through an information-theoretic lens (Zhang et al., 8 May 2026). Under the information bottleneck principle, it maximizes view-invariant information and compresses view-specific noise. Its Cross-View Adaptive Concept Selection filters view-specific slots, while Concept Structural Relational Reasoning aligns object-centric structural relations via graph Laplacians and an Orthogonal Procrustes loss.

These lines are complementary rather than mutually exclusive. One family repairs noisy pair supervision, another purifies sparse correspondences, another suppresses irrelevant attention responses, and another regularizes representations to preserve only shared cross-view structure.

4. PAUL and the explicit NC-CVGL framework

PAUL is the first method in the provided literature to formulate NC-CVGL explicitly and to build an end-to-end learning framework around it (Li et al., 27 Aug 2025). The framework operates on two peer networks, A and B, and consists of three coupled stages: co-divide, co-augment, and co-training.

In co-divide, each network computes a standard InfoNCE loss for every pair,

$\mathcal{R}$5

where $\mathcal{R}$6 is feature cosine similarity. A two-component Gaussian Mixture Model is fitted over the loss values:

$\mathcal{R}$7

The posterior “cleanness” weight

$\mathcal{R}$8

induces a soft partition into $\mathcal{R}$9 and $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$0.

In co-augment, noisy pairs in $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$1 are processed by Evidential Deep Learning. Similarity logits are converted to evidence and Dirichlet parameters,

$$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$2

from which the predictive mean, predictive variance, and uncertainty

$$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$3

are derived. A batch-wise EDL loss combines an MSE term, a variance term, and a KL regularizer toward $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$4. PAUL then computes a Grad-CAM style saliency map on the EDL loss, thresholds it at percentile $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$5, retains the largest connected component, and masks low-confidence pixels in both query and reference. The resulting augmented pairs form $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$6.

In co-training, network A trains on $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$7, and network B trains symmetrically on A’s partitions. For each peer network $$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$8,

$$IoU(q_i,r_j)=\frac{\mathrm{area}(q_i\cap r_j)}{\mathrm{area}(q_i\cup r_j)}.$$9

$P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$0

and

$P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$1

Architecturally, PAUL uses a ViT-Base backbone for both UAV and satellite images at input resolution $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$2, with no specialized uncertainty head; uncertainty is estimated directly from batch-wise similarity logits. Training begins with a one-epoch InfoNCE warm-up, followed by alternating partition, augmentation, exchange, and update steps.

A central conceptual point is that PAUL does not treat noisy samples only as candidates for filtering or label correction. It uses both data uncertainty and loss discrepancy for targeted partitioning and augmentation, thereby extracting useful supervision from semi-positive pairs.

5. Benchmarks, metrics, and empirical behavior

The empirical literature on NC-CVGL and related noise-robust CVGL uses heterogeneous evaluation protocols. PAUL reports Recall@K, Average Precision, SDM@3, and top-1 localization error Dis@1 on same-area and cross-area splits (Li et al., 27 Aug 2025). The uncertainty-aware metric CVGL model reports Mean Absolute Error, RMSE, and lateral/longitudinal recall (Fervers et al., 2022). View Consistent Purification reports median/mean translation and yaw errors together with Recall@$P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$3 and Recall@$P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$4 (Wang et al., 2023). AttenGeo uses [email protected] and [email protected] (Zhu, 31 Oct 2025), whereas InfoGeo emphasizes Recall@1 and AP across transfer scenarios (Zhang et al., 8 May 2026).

PAUL is evaluated on GTA-UAV and UAV-VisLoc with synthetic noise ratios of $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$5, $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$6, and $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$7 generated by replacing aligned pairs with semi-positive pairs (Li et al., 27 Aug 2025). On GTA-UAV cross-area, PAUL achieves $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$8 versus InfoNCE $P=\{(q_i,r_j)\mid IoU(q_i,r_j)>\tau_m\},$9 at $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$0 noise, with lowest $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$1 m. At $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$2 noise, PAUL reaches $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$3 versus $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$4, and at $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$5 noise it reaches $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$6 versus $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$7. On real-world UAV-VisLoc at $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$8 noise, PAUL reports $N=\{(q_i,r_j)\mid \tau_s<IoU(q_i,r_j)\le \tau_m\},$9 versus InfoNCE $\tau_m=0.39$0, with AP and Dis@1 also improved. Ablation studies show that removing either $\tau_m=0.39$1 or $\tau_m=0.39$2 degrades $\tau_m=0.39$3 by approximately $\tau_m=0.39$4–$\tau_m=0.39$5 points under $\tau_m=0.39$6 noise.

The uncertainty-aware metric CVGL model demonstrates that accurate metric localization remains feasible even when ground-truth poses are noisy relative to aerial imagery (Fervers et al., 2022). On Ford AV in the cross-area and cross-vehicle regime, it achieves more than $\tau_m=0.39$7 recall within $\tau_m=0.39$8 m lateral and longitudinal, versus less than $\tau_m=0.39$9 for earlier purely vision methods. When integrated into an IMU/Kalman filter on KITTI-360 using $\tau_s=0.14$0 as measurement weights, it yields a mean position error of $\tau_s=0.14$1 m.

View Consistent Purification reports median lateral and longitudinal accuracy errors below $\tau_s=0.14$2 meters and a median orientation accuracy error below $\tau_s=0.14$3 degrees on KITTI and Ford Multi-AV Seasonal (Wang et al., 2023). On KITTI-CVL, the method achieves median lateral and longitudinal errors of approximately $\tau_s=0.14$4 m and $\tau_s=0.14$5 m, and median yaw of approximately $\tau_s=0.14$6. On Ford Multi-AV Seasonal, median errors remain below $\tau_s=0.14$7 m lateral, below $\tau_s=0.14$8 m longitudinal, and below $\tau_s=0.14$9 yaw, including cross-season and unseen-route settings.

AttenGeo’s experiments show that suppressing edge noise improves cross-view object geo-localization across Ground$N$0Satellite, Drone$N$1Satellite, and the newly introduced Ground$N$2Drone task (Zhu, 31 Oct 2025). On CVOGL test, AttenGeo improves Ground$N$3Satellite from DetGeo $N$4 to $N$5 in [email protected] / [email protected], and Drone$N$6Satellite from $N$7 to $N$8. On G2D test, Ground$N$9Drone improves from $y_{ij}\in\{0,1\}$0 to $y_{ij}\in\{0,1\}$1. The ablations indicate that CVCAM alone improves performance and that MHSAM further sharpens localization.

InfoGeo reports robust transfer performance across University-1652, SUES-200, DenseUAV, and GTA-V cross-area (Zhang et al., 8 May 2026). On University-1652$y_{ij}\in\{0,1\}$2SUES-200, averaged over $y_{ij}\in\{0,1\}$3–$y_{ij}\in\{0,1\}$4 m, InfoGeo with relational distillation reaches $y_{ij}\in\{0,1\}$5 and $y_{ij}\in\{0,1\}$6, compared with CVcities at $y_{ij}\in\{0,1\}$7 and $y_{ij}\in\{0,1\}$8. On DenseUAV$y_{ij}\in\{0,1\}$9SUES-200, InfoGeo without relational distillation achieves $\mathcal{Q}$00 and $\mathcal{Q}$01. Under multi-weather Fog$\mathcal{Q}$02Snow on University$\mathcal{Q}$03SUES-200 at $\mathcal{Q}$04 m, InfoGeo reaches $\mathcal{Q}$05 and $\mathcal{Q}$06, compared with CVcities at $\mathcal{Q}$07 and $\mathcal{Q}$08. On GTA-V cross-area, Recall@1 rises from $\mathcal{Q}$09 for CVcities to $\mathcal{Q}$10 for InfoGeo and $\mathcal{Q}$11 for InfoGeo*.

6. Limitations, distinctions, and open directions

The current NC-CVGL literature also makes clear that robustness remains conditional on the noise model used. PAUL’s noisy supervision is semi-synthetic, derived from partial-overlap samples in GTA-UAV, and the authors explicitly identify the need for larger real-world noisy datasets and online noise detection on live UAV platforms (Li et al., 27 Aug 2025). AttenGeo notes a portability limitation: when CVCAM replaces DetGeo’s original fusion in isolation, it improves Ground$\mathcal{Q}$12Satellite and Ground$\mathcal{Q}$13Drone but slightly degrades Drone$\mathcal{Q}$14Satellite, with visualizations suggesting that CVCAM can re-activate non-target regions when cross-view geometry is too ambiguous (Zhu, 31 Oct 2025). The G2D dataset fills the Ground$\mathcal{Q}$15Drone gap, but broader coverage over altitude, rural scenes, and weather is still identified as necessary.

A second distinction concerns what is being corrected. In uncertainty-aware metric CVGL, the main issue is inaccurate geo-pose supervision, and the remedy is pseudo-label graph optimization and uncertainty-aware tracking (Fervers et al., 2022). In View Consistent Purification, the core issue is corrupted local correspondence due to dynamic objects and non-ground geometry, and the remedy is confidence-based keypoint filtering (Wang et al., 2023). In PAUL, the issue is partial overlap in nominal positives, and the remedy is uncertainty-guided partition and augmentation (Li et al., 27 Aug 2025). In InfoGeo, the objective is to compress view-specific noisy signals while preserving object-centric structural relations (Zhang et al., 8 May 2026).

A plausible implication is that future NC-CVGL systems will combine these mechanisms: uncertainty-aware pair partitioning, explicit geometric priors, object-centric structural reasoning, and temporal filtering. That implication is consistent with the directions already proposed in the literature, including explicit scene-structure or altitude cues, dynamic iteration counts in CVCAM, stronger regularizers against residual noise leakage, larger real-world noisy datasets, and online noise detection (Zhu, 31 Oct 2025). Under that broader interpretation, NC-CVGL has become a unifying term for robustness-oriented cross-view geo-localization in the presence of partial overlap and view-dependent nuisance structure.