CVUSA: Cross-View Geo-localization Benchmark

• CVUSA is a large-scale dataset featuring 35,532 training and 8,884 test pairs of 360° street-view panoramas and north-aligned satellite images for geo-localization and synthesis tasks.
• It uses a center-aligned, north-oriented protocol that serves as a canonical benchmark for standard retrieval, orientation estimation, and limited field-of-view evaluations.
• Recent methods on CVUSA achieve near-saturated Recall@K scores while exploring realistic conditions and cross-view synthesis techniques to bridge domain gaps.

Cross-View USA (CVUSA) is a large-scale cross-view dataset of geographically paired street-view panoramas and overhead satellite images from the United States, and it has become one of the canonical benchmarks for cross-view geo-localization and cross-view image synthesis. In the standard protocol used across much of the literature, CVUSA provides 35,532 training pairs and 8,884 test pairs, with a one-to-one correspondence between each ground panorama and its matching satellite tile; the canonical task is ground-to-satellite retrieval under north-aligned, center-aligned conditions, typically evaluated with Recall@K metrics such as $R@1$, $R@5$, $R@10$, and $R@1\%$ (Li et al., 17 May 2025, Xia et al., 2024, Lin et al., 2024).

1. Dataset definition and canonical benchmark protocol

CVUSA was originally introduced by Zhai et al. to study ground-level layout prediction from aerial imagery, and later became a standard benchmark for cross-view retrieval, orientation estimation, and synthesis (Lin et al., 2024). In the form used by many later papers, it contains 44,416 paired samples split into 35,532 training pairs and 8,884 test pairs, with each pair consisting of a ground-level panoramic image and a north-aligned overhead satellite image from the same geographic location (Toker et al., 2021, Shi et al., 2019).

Several works make the raw modality assumptions explicit. The ground view is a full $360^\circ$ panorama, commonly reported at $1232 \times 224$, while the satellite image is a square tile, commonly reported at $750 \times 750$ (Deuser et al., 2023, Bajbaa et al., 29 Sep 2025). The satellite patch is centered at the ground camera location, and the top of the aerial image corresponds to geographic north; in parallel, the panorama is treated as north-aligned in the standard setup (Shi et al., 2019, Toker et al., 2021). This center-aligned, north-aligned design makes CVUSA a particularly clean benchmark for studying viewpoint change without simultaneously introducing large position offsets.

The standard evaluation protocol is image retrieval. A ground panorama is used as query, all test-set satellite images form the gallery, and the task is to rank the correct aerial tile as highly as possible. One formulation used on CVUSA writes cross-view retrieval as

$$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$

where $sim(\cdot,\cdot)$ is cosine similarity between the query ground embedding and gallery aerial embeddings (Xia et al., 2024). In most retrieval papers using CVUSA, the correspondence is strictly one-to-one and performance is summarized by Recall@K rather than metric-distance thresholds (Shi et al., 2019, Li et al., 17 May 2025).

Although the raw dataset is fixed, input resolutions are highly method-dependent. Examples include $140 \times 768$ ground and $R@5$0 satellite for Sample4Geo-style retrieval training, $R@5$1 for polar-transformed satellite-to-street multi-task learning, $R@5$2 panoramas with $R@5$3 satellite patches for diffusion-based synthesis, and $R@5$4 crops for GAN-based translation (Deuser et al., 2023, Toker et al., 2021, Lin et al., 2024, Tang et al., 2019). These resizings are task-specific preprocessing choices rather than alternative definitions of the benchmark.

2. Geometric assumptions, alignment regimes, and benchmark variants

A defining property of CVUSA is that it is a center-aligned benchmark with no decentrality. In the decentrality study of AuxGeo and DReSS, CVUSA is explicitly treated as the canonical “no-decentrality” case: the query location matches the center of the aerial tile, no positional offset is simulated, and all decentrality analysis is moved to other datasets such as CVSat and VIGOR (Xia et al., 2024). This makes CVUSA useful as a controlled baseline, but it also means that very high retrieval scores on CVUSA do not, by themselves, establish robustness to real-world miscentering.

The standard CVUSA protocol also assumes known orientation in the sense that ground panoramas and aerial images are north-aligned. A substantial body of work therefore modifies CVUSA to study more realistic conditions. DSM converts the north-aligned aerial image into a polar representation and then evaluates not only the original $R@5$5 panorama setting, but also panoramas with unknown orientation and limited-field-of-view crops at $R@5$6, $R@5$7, and $R@5$8, obtained by randomly shifting and cropping the panorama horizontally (Shi et al., 2020). Fine-grained orientation work follows the same idea, creating random orientation shifts and then measuring angular error directly; on CVUSA, this yields $R@5$9 accuracy for estimated angle errors below $R@10$0 with the best variant (Hu et al., 2023).

A second line of work adapts CVUSA to realistic forward-facing camera regimes. BEV-CV uses the yaw metadata to crop $R@10$1 and $R@10$2 limited-FoV views aligned to vehicle direction, arguing that this better matches monocular vehicle cameras than full panoramas do (Shore et al., 2023). W2W-BEV likewise randomizes panorama orientation and then crops limited-FoV windows before resizing them for retrieval, explicitly targeting the regime of unknown orientation and limited FoV (Cheng et al., 2024). These variants convert CVUSA from a clean aligned benchmark into a controlled testbed for orientation ambiguity and partial observability.

This multiplicity of protocols is important for interpretation. “CVUSA performance” can refer to at least four distinct regimes: standard center-aligned panoramic retrieval, unknown-orientation panoramic retrieval, limited-FoV retrieval, and cross-view image synthesis in either direction. The benchmark is therefore more heterogeneous in practice than its canonical retrieval setup might suggest.

3. Methodological roles in cross-view geo-localization

CVUSA has been central to several successive methodological paradigms in cross-view retrieval. Early work emphasized explicit handling of the domain gap between ground panoramas and aerial imagery. “Bridging the Domain Gap for Ground-to-Aerial Image Matching” synthesizes an aerial representation from a ground panorama with a conditional GAN, then uses joint feature learning and feature fusion to reduce the gap between views (Regmi et al., 2019). “Optimal Feature Transport for Cross-View Image Geo-Localization” instead keeps spatial feature maps and learns a differentiable transport matrix via entropy-regularized optimal transport and Sinkhorn normalization, improving CVUSA top-1 recall from 40.79% to 61.43% relative to a strong earlier baseline (Shi et al., 2019).

A second strand focused on orientation and part structure. DSM applies a polar transform to aerial images and computes a cross-correlation over azimuth to jointly estimate location and orientation, leading to 91.96% $R@10$3 on standard CVUSA and much larger gains under unknown orientation and limited FoV (Shi et al., 2020). “Each Part Matters” introduces Local Pattern Network (LPN), which partitions features part-wise and uses contextual neighbor areas as auxiliary information, achieving 85.79% $R@10$4 on CVUSA with a ResNet-50 backbone and improving both SAFA and CVFT when plugged into them (Wang et al., 2020). “Coming Down to Earth” couples retrieval and satellite-to-street synthesis end-to-end, sharing the generator encoder with the retrieval branch and reporting 92.56% $R@10$5 on CVUSA (Toker et al., 2021).

More recent work has shifted toward contrastive learning, BEV intermediaries, and explicit cross-view correspondence. Sample4Geo shows that a shared ConvNeXt-B encoder with symmetric InfoNCE and hard negative sampling can outperform more elaborate polar-transform or aggregation-heavy pipelines, reaching 98.68% $R@10$6 on CVUSA without polar preprocessing (Deuser et al., 2023). BEV-CV and the Panorama-BEV Co-Retrieval Network both reduce the viewpoint gap by converting panoramas into bird’s-eye-view representations; the former concentrates on realistic limited-FoV inputs, while the latter retains both panorama and BEV branches and achieves 98.71% $R@10$7 in the standard centered protocol (Shore et al., 2023, Ye et al., 2024). W2W-BEV goes further by learning a BEV representation directly from limited-FoV ground images and matching BEV windows to ground windows under unknown orientation (Cheng et al., 2024).

The latest CVUSA results are increasingly dominated by training-time auxiliary structure rather than inference-time complexity. AuxGeo adds the Bird’s-eye view Intermediary Module (BIM) and Position Constraint Module (PCM), both used only during training, and still improves Sample4Geo in the standard centered setting to 98.80% $R@10$8 (Xia et al., 2024). CVD explicitly disentangles content and viewpoint, performs retrieval only in the content space, and improves multiple baselines on CVUSA, including Sample4Geo from 98.43% to 98.67% $R@10$9 (Li et al., 17 May 2025). CLNet introduces view neural maps, a Neural Bird’s-Eye View Converter, and Global Feature Recalibration, reaching 98.77% $R@1\%$0 and 99.89% $R@1\%$1 (Cao et al., 16 Dec 2025). Geo$R@1\%$2 uses VGGT-derived geometric priors to build a shared 3D-aware latent space and reports 98.83% $R@1\%$3, while also unifying localization with synthesis (Zhang et al., 26 Mar 2026).

A recurring methodological theme is that CVUSA rewards models that exploit global structure plus local spatial correspondence. The exact mechanism varies—polar transforms, optimal transport, BEV projection, auxiliary geometry losses, disentanglement, or frozen geometric foundation models—but the benchmark consistently favors approaches that narrow the appearance and topology gap between panoramic ground imagery and overhead satellite imagery.

4. Cross-view image synthesis on CVUSA

CVUSA is also a major benchmark for cross-view image synthesis, where the task is to generate a ground panorama from satellite imagery or vice versa. Because the benchmark pairs images at the same location under extreme viewpoint change, it supports both conditional generation and retrieval-aware evaluation of generated images.

One of the early synthesis uses of CVUSA is “Multi-Channel Attention Selection GAN with Cascaded Semantic Guidance,” which evaluates aerial-to-ground translation on CVUSA using a cropped quarter of the panorama and an automatically generated target semantic map. SelectionGAN is a two-stage system: a coarse semantic-guided generator followed by a multi-channel attention refinement module with uncertainty-guided pixel loss (Tang et al., 2019). On CVUSA aerial-to-ground synthesis it reports SSIM 0.5323, PSNR 23.15, and KL 2.96, outperforming Zhai et al., Pix2pix, X-Fork, and X-Seq under that protocol (Tang et al., 2019).

A separate line ties synthesis directly to localization. “Retrieval-guided Cross-view Image Synthesis” conditions both generator and discriminator on embeddings from a frozen SAIG retrieval model, using Attn-AdaIN to inject view-invariant structure and style codes for view-specific details (Yang et al., 2024). On CVUSA, this method reports for aerial-to-ground synthesis SSIM 0.3706, PSNR 14.33, LPIPS 0.4302, FID 13.57, and retrieval-based $R@1\%$4 96.25; for ground-to-aerial synthesis it reports SSIM 0.2461, PSNR 15.77, LPIPS 0.5181, FID 41.65, and $R@1\%$5 95.14 (Yang et al., 2024). The unusually high retrieval-based $R@1\%$6 scores indicate that generated images preserve location identity in a strong cross-view embedding space.

Diffusion-based CVUSA synthesis has developed along two main axes. “Geometry-guided Cross-view Diffusion for One-to-many Cross-view Image Synthesis” treats both satellite-to-ground and ground-to-satellite generation as one-to-many problems and introduces the Geometry-guided Cross-view Condition (GCC), which projects multi-level conditioning features into the target-view coordinate system before diffusion denoising (Lin et al., 2024). On CVUSA satellite-to-ground synthesis, the ControlNet variant reports PSNR 14.274, LPIPS$R@1\%$7 0.4345, and FID 13.755; on ground-to-satellite synthesis, it reports PSNR 14.070, SSIM 0.2271, and FID 53.080 (Lin et al., 2024). “From Satellite to Street” later combines Stable Diffusion v2.1, ControlNet, PanoGAN, and a fusion network in a hybrid panorama-synthesis pipeline evaluated exclusively on CVUSA (Bajbaa et al., 29 Sep 2025).

Geo$R@1\%$8 places CVUSA synthesis inside a joint localization-and-synthesis framework. It uses a geometry-aware latent space from GeoMap and a reversible flow-matching model, GeoFlow, for bidirectional generation (Zhang et al., 26 Mar 2026). On CVUSA ground-to-satellite synthesis it reports LPIPS 0.534 and SSIM 0.167, while on satellite-to-ground synthesis it reports FID 29.05 and PSNR 12.94 (Zhang et al., 26 Mar 2026). A plausible implication is that CVUSA has become not only a retrieval benchmark but also a testbed for whether a single geometry-aware representation can support both recognition and generation.

Metric comparability across synthesis papers is not perfect. Some protocols operate on full panoramas, others on $R@1\%$9 panoramas, $360^\circ$0 panoramas, or only the first quarter of the panorama; some report averages over multiple diffusion samples, while others evaluate deterministic GAN outputs (Tang et al., 2019, Lin et al., 2024, Bajbaa et al., 29 Sep 2025). As a result, synthesis numbers on CVUSA are best interpreted within each protocol family rather than as a single totally uniform leaderboard.

5. Performance trajectory and representative results

The retrieval literature on CVUSA shows a clear trajectory from domain-gap handling and orientation modeling to near-saturated top-1 recall in the canonical centered protocol.

Method	CVUSA setting	Representative result
CVFT (Shi et al., 2019)	Standard retrieval	$360^\circ$1, $360^\circ$2, $360^\circ$3
DSM (Shi et al., 2020)	Panoramas, known orientation	$360^\circ$4
CDE (Toker et al., 2021)	Joint retrieval and synthesis	$360^\circ$5
Sample4Geo (Deuser et al., 2023)	Standard centered retrieval	$360^\circ$6, $360^\circ$7, $360^\circ$8
Panorama-BEV Co-Retrieval (Ye et al., 2024)	Standard centered retrieval	$360^\circ$9, $1232 \times 224$0
AuxGeo (Xia et al., 2024)	Standard centered retrieval	$1232 \times 224$1, $1232 \times 224$2
CLNet (Cao et al., 16 Dec 2025)	Standard centered retrieval	$1232 \times 224$3, $1232 \times 224$4
Geo$1232 \times 224$5 (Zhang et al., 26 Mar 2026)	Standard centered retrieval	$1232 \times 224$6, $1232 \times 224$7

Under more realistic orientation or FoV settings, the gains are less saturated and therefore more revealing. DSM raises CVUSA $1232 \times 224$8 under unknown-orientation panoramas to 78.11%, and on $1232 \times 224$9 FoV unknown-orientation crops to 48.53% (Shi et al., 2020). Fine-grained orientation estimation later reaches $750 \times 750$0 accuracy within $750 \times 750$1 on CVUSA while also improving geo-localization to 95.50% $750 \times 750$2 in the orientation-known case and 85.45% in the orientation-unknown case (Hu et al., 2023). BEV-CV reports 14.03% $750 \times 750$3 on CVUSA $750 \times 750$4 crops under its standard protocol and 27.40% under its orientation-aware protocol, emphasizing the difficulty of realistic forward-facing inputs (Shore et al., 2023). W2W-BEV reports the largest limited-FoV jump: on CVUSA with FoV $750 \times 750$5 and unknown orientation, $750 \times 750$6 rises from 47.24% for Sample4Geo to 64.75% (Cheng et al., 2024).

Synthesis performance has evolved in parallel.

Method	CVUSA synthesis task	Representative result
SelectionGAN (Tang et al., 2019)	Aerial $750 \times 750$7 ground	SSIM 0.5323, PSNR 23.15, KL 2.96
Retrieval-guided GAN (Yang et al., 2024)	Aerial $750 \times 750$8 ground	SSIM 0.3706, FID 13.57, retrieval $750 \times 750$9
Retrieval-guided GAN (Yang et al., 2024)	Ground $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$0 aerial	SSIM 0.2461, FID 41.65, retrieval $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$1
GCC diffusion (Lin et al., 2024)	Satellite $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$2 ground	PSNR 14.274, FID 13.755
Geo$$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$3 (Zhang et al., 26 Mar 2026)	Ground $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$4 satellite	LPIPS 0.534, SSIM 0.167

These numbers should not be collapsed into a single ranking, because CVUSA synthesis papers differ in panorama coverage, resolution, determinism versus stochasticity, and evaluation protocol. They do, however, show that CVUSA has supported both retrieval-oriented and perceptual notions of progress.

6. Benchmark status, limitations, and continuing relevance

A persistent misconception is that strong CVUSA performance necessarily implies strong real-world geo-localization. The decentrality analysis in AuxGeo argues directly against this reading: CVUSA has no decentrality, the query is centered in the aerial tile, and the benchmark is therefore a clean but idealized test bed (Xia et al., 2024). The same work further characterizes CVUSA as suburban and relatively easier than decentrality-oriented datasets, with top-1 recall near 99% indicating a saturated regime (Xia et al., 2024).

A second misconception is that CVUSA is only a retrieval dataset. In practice it has become a benchmark for at least four technically distinct problems: standard ground-to-aerial retrieval, orientation estimation under unknown yaw, limited-FoV geo-localization, and cross-view image synthesis in both directions (Shi et al., 2020, Hu et al., 2023, Yang et al., 2024). This breadth is one reason it remains prominent despite saturation in its original centered protocol.

The main limitation of CVUSA is therefore not lack of utility, but lack of realism in some canonical assumptions. Center alignment, fixed north orientation, one-to-one correspondences, and strong panorama coverage simplify the retrieval problem relative to large-area GNSS-denied deployment (Xia et al., 2024, Shore et al., 2023). Several papers explicitly respond by creating CVUSA-derived realistic settings—unknown orientation, $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$5 or $$r = \arg\max_{i \in \{1,\dots,N\}} sim(f_q^s, f_i^a),$$6 FoV crops, heading-aware protocols—or by treating CVUSA as the baseline “easy” case against which harder datasets should be compared (Shi et al., 2020, Shore et al., 2023, Cheng et al., 2024).

For that reason, CVUSA now occupies a dual position in the literature. It is still the canonical benchmark for controlled cross-view matching under clean alignment, and it remains a standard site for demonstrating incremental gains from new losses, BEV representations, disentanglement modules, or geometric priors. At the same time, its very cleanliness has made it a reference point for what newer benchmarks must move beyond. The benchmark’s enduring importance lies less in its ability to expose every real-world difficulty than in its role as the shared baseline through which different generations of cross-view geo-localization and synthesis methods can still be compared.