Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming

Yunus Talha Erzurumlu; Jiyong Kwag; Alper Yilmaz

cs.CVMar 26, 2026

Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming

Authors: Yunus Talha Erzurumlu, Jiyong Kwag, Alper Yilmaz

Organizations: Photogrammetric Computer Vision Lab, The Ohio State University

Abstract

Cross-view geo-localization (CVGL) estimates a camera's location by matching a street-view image to geo-referenced overhead imagery, enabling GPS-denied localization and navigation. Existing methods almost universally formulate CVGL as an image-retrieval problem in a contrastively trained embedding space. This ties performance to large batches and hard negative mining, and it ignores both the geometric structure of maps and the coverage mismatch between street-view and overhead imagery. In particular, salient landmarks visible from the street view can fall outside a fixed satellite crop, making retrieval targets ambiguous and limiting explicit spatial inference over the map. We propose Just Zoom In, an alternative formulation that performs CVGL via autoregressive zooming over a city-scale overhead map. Starting from a coarse satellite view, the model takes a short sequence of zoom-in decisions to select a terminal satellite cell at a target resolution, without contrastive losses or hard negative mining. We further introduce a realistic benchmark with crowd-sourced street views and high-resolution satellite imagery that reflects real capture conditions. On this benchmark, Just Zoom In achieves state-of-the-art performance, improving Recall@1 within 50 m by 5.5% and Recall@1 within 100 m by 9.6% over the strongest contrastive-retrieval baseline. These results demonstrate the effectiveness of sequential coarse-to-fine spatial reasoning for cross-view geo-localization.

Figures & tables

Figure 1: Two paradigms for cross-view geo-localization. (a) Retrieval-based approaches maintain a large GPS-tagged reference database and use a contrastively trained encoder to perform nearest-neighbor search at query time. (b) Autoregressive geo-localization (ours) replaces global retrieval with sequential, coarse-to-fine zoom-in decisions over multi-scale satellite imagery, dramatically reducing dependence on exhaustive database search.

Figure 2: Coverage mismatch. In the street-view image (a), the stadium and its distinctive gate are clearly visible. In the satellite crop (b) from the same region, this landmark falls outside of the patch, illustrating how small tiles can miss critical cues present in the street-view. Images sourced from Google Street View and Google Maps [ 8 , 9 ] .

Dataset	# Photos	Multi-Scale	Ground / Pairing Protocol			Domain Types
			SV GPS	Gnd FoV	Heading	Urban	Suburban	Rural	Offroad
CVUSA [ 39 ]	44K	✗	Aligned	360° pano	Fixed	✓	✓	✓	✓
CVACT [ 19 ]	128K	✗	Aligned	360° pano	Fixed	✓	✓	✗	✗
Vo & Hayes [ 38 ]	450K	✗	Aligned	Crops	Arbitrary	✓	✓	✗	✗
VIGOR [ 44 ]	105K	✗	Arbitrary	360° pano	Fixed	✓	✗	✗	✗
Ours	300K	✓	Arbitrary	Perspective	Arbitrary	✓	✓	✓	✓

Table 1: Dataset comparisons for cross-view geo-localization. We compare our proposed multi-scale benchmark against prior datasets under key capture and supervision attributes. SV GPS : Aligned = center-aligned one-to-one pairing; Arbitrary = query location not constrained to the tile center. Gnd FoV : 360° pano = panorama queries; Crops = crops from panoramas; Perspective = native non-panoramic images. Heading : Fixed = known/normalized orientation; Arbitrary = unknown or random orientation.

Figure 3: Dataset examples. Street-view samples from our cross-view image localization corpus illustrating the two defining characteristics: (i) limited FoV typical of first-person and dash-cam capture; and (ii) broad diversity across time of day, seasonal appearance, weather, scene type , and capture platforms . This variability-together with viewpoint changes, occlusions, and motion blur-widens the appearance gap to overhead imagery and makes cross-view matching more challenging.

Figure 4: Overview of Just Zoom In . (a) Image encoding. A shared, off-the-shelf vision encoder maps the street-view image $I_{g}$ and the multi-scale satellite maps $\{M_{t}\}$ to global representation tokens $e_{g}$ and $e_{t}$ , single token per image. (b) Autoregressive action modeling. The image tokens, interleaved with previously chosen action tokens, are fed to a causal transformer that predicts a distribution over the next zoom-in action $a_{t}$ at each step. (c) Sample / zoom-in sequence. Beginning from $M_{0}$ , the model selects the most probable patch, zooms to obtain $M_{t+1}$ , and iterates until a terminal patch $M_{N}$ ; the center of $M_{N}$ is taken as the location estimate.

Method	R@40m $\uparrow$	R@50m $\uparrow$	R@100m $\uparrow$	Inf. Mem. (GB) $\downarrow$	Train. Time $\downarrow$	HNM
SAIG-D [ 45 ]	39.36	47.52	64.17	18	$>90$	✓
TransGeo [ 43 ]	45.97	54.55	67.61	11	$>90$	✓
Sample4Geo [ 5 ]	52.85	60.81	71.30	16	$>90$	✓
Ours (BS. = 32)	51.08	61.49	77.43	9	61	✗
Ours (BS. = 64)	54.10	64.74	79.74	9	58	✗
Ours (BS. = 128)	55.74	66.31	80.93	9	52	✗

Table 2: Cross-view geo-localization results on our proposed dataset. We reproduce all contrastive-based baselines using their official implementations. Our coarse-to-fine autoregressive approach outperforms retrieval-based methods across all evaluation thresholds. R@

\tau

m denotes Recall@1 within

\tau

meters. Inf. Mem. (GB) reports the peak memory required to store and search the reference database at inference time, Train. Time (GPU hours) is the approximate end-to-end training time on one NVIDIA A6000 GPU, and HNM indicates whether the method relies on hard negative mining during training.

Figure 5: Focus at different zoom levels. We visualize similarity maps between street-view patch tokens and the satellite [CLS] embedding (blue = low similarity, red = high similarity) across three zoom levels. At the coarsest level, high similarity concentrates on far-away regions, whereas at finer zoom levels it shifts toward nearby objects and road/building details, indicating that the model learns to exploit different scale-specific visual cues.

Method	R@40m $\uparrow$	R@50m $\uparrow$	R@100m $\uparrow$
[CLS] + 256	63.38	68.25	79.83
[CLS] + 64	64.60	70.10	80.70
[CLS] + 32	65.80	72.10	83.02
[CLS] + 16	65.31	71.6	82.60
[CLS] + 8	65.31	70.76	82.12
[CLS] Only	66.98	72.25	83.08

Table 4: Comparison of different satellite image representations. We compare global [CLS] token features with variants that incorporate local patch tokens from the vision encoder. Our results show that relying on the global [CLS] representation yields the highest localization accuracy.

Figure 6: Results. Qualitative visualization of our model’s auto-regressive localization. The first column shows the input street-view query. The subsequent columns (Lvl 1-4) display the satellite map at each sequential zoom step. On each map, the model’s predicted patch is highlighted by the grid overlay, the ground truth (GT) location is marked with a red dot, and the 50m success radius is shown as a white dotted circle.

Cross-View Geo-Localization

Cross-view geo-localization matches ground, drone, or aerial images to maps or satellite views to estimate where they were captured.

See all 94 papers

One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization
Aug 2, 2026Ruijie Fan, Junyan Ye, Qi Zhu +1Cross-View Geo-Localization Geospatial Representation Learning
Tsinghua Shenzhen International Graduate School, Tsinghua University · School of Geospatial Engineering and Science, Sun Yat-sen University
BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization
May 11, 2026Wei Wang, Dou Quan, Ning Huyan +4Cross-View Geo-Localization Geospatial Representation Learning
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University · Department of Automation, Tsinghua University, Beijing 100084, China · School of Telecommunications, Xidian University, Xi’an 710071, China
GeoBridge++: Fact-Guided Geo-Semantic Bridging for Unified Cross-View Geo-Localization
Dec 2, 2025Zixuan Song, Jing Zhang, Di Wang +6Cross-View Geo-Localization Geospatial Representation Learning
College of Computer Science and Technology and Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun 130012, China · Zhongguancun Academy, Beijing 100094, China · School of Computer Science, Wuhan University, Wuhan, China +1

Aug 2, 2026J/K move · Enter open · S save

One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization

Ruijie Fan, Junyan Ye, Qi Zhu, Weijia Li

Tsinghua Shenzhen International Graduate School, Tsinghua University · School of Geospatial Engineering and Science, Sun Yat-sen University

Cross-View Geo-Localization Geospatial Representation Learning Sparse Mixture-of-Experts Remote Sensing Benchmarks

Open paper AskPDF arXiv

Cross-view geo-localization (CVGL) retrieves geo-tagged satellite imagery for a ground-view query. Most systems exhaustively search a flat, fixed-resolution gallery, incurring high cost over large areas and adapting poorly to satellite resolution changes. Autoregressive coarse-to-fine alternatives reduce comparisons but bind later predictions to earlier decisions and a predefined hierarchy. We introduce GeoMoE, a sparse mixture-of-experts dual encoder that decouples global multi-scale representation learning from local hierarchical search. Global multi-scale supervision and content-adaptive routing map ground and satellite images across resolutions into a globally comparable embedding space. At inference, each image is encoded once, and probabilistic beam search follows parent--child links to score a small candidate subset. Later levels reuse these descriptors rather than features generated by preceding levels, limiting feature-level error propagation and hierarchy coupling. We further introduce VIGOR-M, a four-city benchmark with an explicit parent--child satellite hierarchy and held-out half-step galleries for single-resolution, cross-resolution, and hierarchical evaluation. GeoMoE achieves 95.78% R@40m on Just Zoom In, 2.77 percentage points above the previous best, and 62.39% R@1 on VIGOR-M. The latter requires 0.885 MMAC/query for descriptor matching, 5.27% of an exhaustive L3 scan, while exceeding the strongest exhaustive baseline by 3.12 percentage points in R@1. One model trained on L1, L2, and L3 also outperforms a matched dense control across all six galleries and transfers to three withheld resolutions. By decoupling globally trained embeddings from local hierarchical search, GeoMoE jointly improves localization accuracy, search efficiency, and cross-resolution transfer.

Abstract

Cross-View Geo-Localization

Explore similar work

One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization

BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization

GeoBridge++: Fact-Guided Geo-Semantic Bridging for Unified Cross-View Geo-Localization