cs.CVFeb 16, 2026

Wrivinder: Towards Spatial Intelligence for Geo-locating Ground Images onto Satellite Imagery

Authors: Chandrakanth Gudavalli, Tajuddin Manhar Mohammed, Abhay Yadav, Ananth Vishnu Bhaskar, Hardik Prajapati, Cheng Peng, Rama Chellappa, Shivkumar Chandrasekaran, +1 more

Organizations: Mayachitra, Inc. · Johns Hopkins University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Aligning ground-level imagery with geo-registered satellite maps is crucial for mapping, navigation, and situational awareness, yet remains challenging under large viewpoint gaps or when GPS is unreliable. We introduce Wrivinder, a zero-shot, geometry-driven framework that aggregates multiple ground photographs to reconstruct a consistent 3D scene and align it with overhead satellite imagery. Wrivinder combines SfM reconstruction, 3D Gaussian Splatting, semantic grounding, and monocular depth--based metric cues to produce a stable zenith-view rendering that can be directly matched to satellite context for metrically accurate camera geo-localization. To support systematic evaluation of this task, which lacks suitable benchmarks, we also release MC-Sat, a curated dataset linking multi-view ground imagery with geo-registered satellite tiles across diverse outdoor environments. Together, Wrivinder and MC-Sat provide a first comprehensive baseline and testbed for studying geometry-centered cross-view alignment without paired supervision. In zero-shot experiments, Wrivinder achieves sub-30,m geolocation accuracy across both dense and large-area scenes, highlighting the promise of geometry-based aggregation for robust ground-to-satellite localization.

Figures & tables

Figure 1 : Overview of the satellite-to-ground image alignment pipeline. Directly aligning ground images to satellite views is impractical due to large viewpoint and scale differences. Wrivinder aggregates information from multiple ground images to reconstruct a 3D scene, generates a zenith-view rendering, and aligns it to the satellite image using the estimated metric dimensions in meters.
Figure 2 : MC-Sat Dataset Overview, showing scenes from the ULTRAA , VisymScenes , and JHU-Ames datasets. The central image in each tile is a satellite view of the scene, surrounded by corresponding ground images illustrating the diversity of viewpoints and environments.
Dataset#Scenes#ImagesImagery Type
ULTRAA [ 15 ]31,028Ground
VisymScenes [ 40 ]149258KGround
ACC - NVS1 [ 33 ]6148KGround + Airborne
JHU-Ames [ 18 ]11,717Ground + Airborne
Table 1: Overview of ground imagery datasets incorporated into the MC-Sat dataset. MC-Sat integrates subsets of these sources and pairs them with orthorectified satellite imagery. More details about the curated subset are reported in Table 2 .
Figure 3 : Overview of Wrivinder , a zero-shot, training-free pipeline for geo-locating ground images on a geo-registered satellite map. Given an unordered set of ground images, the pipeline reconstructs a sparse 3D scene via SfM and densifies it using 3D Gaussian Splatting . The Zenith Viewpoint Extractor estimates the vertical direction and generates a top-down zenith render. The Metric Mapper uses monocular depth priors to recover approximate metric scale and determine the physical footprint of the zenith view. A test-time Deep Template Matcher (DTM) aligns this render to the satellite image, and the resulting correspondences are back-projected through the 3DGS and SfM models via the Gaussian Splat Geolocator to estimate GPS positions for all ground cameras.
Figure 4 : Key intermediate outputs of Wrivinder, showing semantic maps, the SfM point cloud, semantified reconstruction, metric depth maps, and the resulting metric-scaled zenith render.
       Geolocation Error (in meters)
MC-Sat Scene NameDataset TypeSource DatasetSatellite SourceImage CountRun Time (in mins)World2Model RMSEGeolocation RMSE (67th Percentile)Geolocation RMSE (Mean)Geolocation Centroid Error
APL Front DoorImage DensityULTRAANAIP1002280.961.861.960.86
APL Back DoorImage DensityULTRAANAIP1002961.132.562.820.76
MUTC A09Reconstructed AreaULTRAAESRI3344843.3618.3318.8617.34
MUTC A10Reconstructed AreaULTRAAESRI27152215.7617.5917.8216.96
siteSTR0001 (South America)Reconstructed AreaVisymScenesESRI27051560NaN56.8857.2243.82
Table 2 : Quantitative results on the MC-Sat benchmark. We report the number of images, run time, SfM alignment quality (World2Model RMSE), and final geo-localization error across all evaluated scenes. Lower values indicate better alignment.
Figure 5 : Satellite–render pairs for several MC-Sat scenes. In each case, the left image shows the satellite view (with blue dots indicating the ground-truth camera locations) and the right image shows the corresponding 3DGS zenith rendering produced by Wrivinder. Gaps, blurring, and missing structures in the reconstruction often make alignment more ambiguous and are a primary source of the higher errors observed in some scenes.

Explore similar work

CardsList

  1. Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization

    Jun 29, 2026Liyao Wang, Ruipu Wu, Haojun Xu +3Cross-View Geo-LocalizationCross-View

    Beihang University, Beijing, China · Meituan, Beijing, China

  2. Pinpoint: Grounded Worldwide Image Geolocation via Cross-Source Retrieval and Reranking

    Jun 2, 2026Nika Chuzhoy, Brian Hu, Amit A. Arora +2Cross-View Geo-LocalizationVisual Place Recognition

    Virtualitics

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

    Dec 2, 2025Zixuan Song, Jing Zhang, Di Wang +6Cross-View Geo-LocalizationGeogs-Slam

    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