Source-linked AI summary
University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization
Zhedong Zheng, Yunchao Wei, Yi Yang
TL;DR
Cross-view geo-localization needs features robust to large viewpoint changes, but existing benchmarks often provide only two-view image pairs. University-1652 introduces a three-platform benchmark and CNN baseline for two drone-based retrieval tasks; experiments show improved viewpoint-invariant learning and real-world generalization, with campus-building selection limiting landmark-related bias.
Problem
Existing datasets usually provide image pairs from two viewpoints, limiting the multi-view evidence available for learning viewpoint-invariant features.
Method
The paper builds University-1652 from synthetic drone, satellite, and ground-camera images and evaluates CNN retrieval baselines, including a three-branch model.
Results
Experiments show that adding drone viewpoints improves Rank@1 and AP, while learned features generalize to real drone-view images.
Takeaways & Limitations
University-1652 supports drone-view target localization and drone navigation while providing evidence for viewpoint-invariant features and real-world scalability.
Takeaways & Limitations
The dataset uses campus buildings rather than landmarks because landmarks may introduce architectural-style bias and are often restricted for drone flights.
Abstract
from arXiv · showhide
We consider the problem of cross-view geo-localization. The primary challenge of this task is to learn the robust feature against large viewpoint changes. Existing benchmarks can help, but are limited in the number of viewpoints. Image pairs, containing two viewpoints, e.g., satellite and ground, are usually provided, which may compromise the feature learning. Besides phone cameras and satellites, in this paper, we argue that drones could serve as the third platform to deal with the geo-localization problem. In contrast to the traditional ground-view images, drone-view images meet fewer obstacles, e.g., trees, and could provide a comprehensive view when flying around the target place. To verify the effectiveness of the drone platform, we introduce a new multi-view multi-source benchmark for drone-based geo-localization, named University-1652. University-1652 contains data from three platforms, i.e., synthetic drones, satellites and ground cameras of 1,652 university buildings around the world. To our knowledge, University-1652 is the first drone-based geo-localization dataset and enables two new tasks, i.e., drone-view target localization and drone navigation. As the name implies, drone-view target localization intends to predict the location of the target place via drone-view images. On the other hand, given a satellite-view query image, drone navigation is to drive the drone to the area of interest in the query. We use this dataset to analyze a variety of off-the-shelf CNN features and propose a strong CNN baseline on this challenging dataset. The experiments show that University-1652 helps the model to learn the viewpoint-invariant features and also has good generalization ability in the real-world scenario.
1 INTRODUCTION
Cross-view geo-localization requires representations that remain discriminative across large viewpoint and appearance changes. University-1652 addresses the limited-viewpoint setting with multi-view, multi-source data centered on synthetic drone imagery and evaluates two drone-based tasks.
- Existing datasets mainly provide ground–satellite image pairs, which may compromise learning viewpoint-invariant features.
- Drones can capture comprehensive target views with fewer obstacles while flying around buildings, complementing ground and satellite imagery.
- University-1652 contains 54 synthetic drone-view images per place, alongside satellite and ground-view images, across 1,652 buildings from 72 universities.
- The benchmark introduces drone-view target localization, matching drone queries to satellite images, and drone navigation, matching satellite queries to previously visited drone-view locations.
- Experiments compare generic CNN features, baseline architectures, and loss functions, including tests on real drone-view images for scalability.
2 RELATED WORK
Prior geo-localization work commonly treats cross-view matching as image retrieval and learns shared representations across limited pairs of platforms. The paper’s multi-view framing extends this background with drone and ground-view collection perspectives.
- Previous geo-localization datasets generally match image pairs from two platforms, such as phone cameras and satellites.
- Figure 2 depicts drone flight around a target and street-view collection from multiple building facets, providing scale, viewpoint, and facade variation.
- A central research goal is learning a viewpoint-invariant representation that bridges visual gaps between images from different views.
3 UNIVERSITY-1652 DATASET
University-1652 combines satellite, synthetic drone, and ground imagery for university buildings, with multi-view references, separated university splits, distractors, and retrieval-oriented evaluation.
- Dataset Description: The dataset selects 1,652 campus buildings from 72 universities rather than landmarks to reduce architectural-style bias and reflect drone-accessible settings.
- Dataset Description: Building metadata is linked to geographic coordinates through university records and map search, excluding ambiguous results.
- Dataset Description: Ground imagery combines manually selected street views with image-search results, while filtering removes indoor and duplicate images.
- Evaluation Protocol: Multiple query images enable a multiple-query setting that produces more accurate target localization in the experiments.
- Evaluation Protocol: The benchmark separates 701 training buildings from 701 test buildings across non-overlapping university sets, with 250 additional buildings used as gallery distractors.
- Evaluation Protocol: Average precision complements Recall@K because the gallery contains multiple true matches from different viewpoints.
4 CROSS-VIEW IMAGE MATCHING
Cross-view matching maps images from different sources into a shared space so same-location embeddings are close and different-location embeddings are separated. The baseline uses source-specific CNN branches, shared classification weights, and instance loss to align satellite, ground, and drone representations.
- Cross-view matching maps images from different sources into a shared space, bringing same-location embeddings close and separating different locations.
- The architecture includes satellite-ground, satellite-drone, and three-branch variants that jointly use all three annotated platforms.
- The method uses separate projection functions and CNN branches because satellite, ground, and drone images have different low-level patterns.
- Instance loss treats each place as one class and uses a shared classifier to map images from different sources into a shared feature space.
- Keeping shared classification weights for extra sources extends the soft feature-alignment constraint to multi-source data.
5 EXPERIMENT
The experiments evaluate cross-view retrieval settings, feature choices, training viewpoints, losses, architectures, input sizes, and transfer performance. Results support multi-view training and the proposed learned features, while also showing task-dependent effects across model designs and datasets.
- 5.2 Geo-localization Results: The evaluation treats drone-view target localization and drone navigation as cross-view image retrieval from source A to gallery B.The experiments compare generic CNN features, learned features, model variants, and training objectives across these settings.
- 5.2 Geo-localization Results: The learned 512-dimensional feature yields better University-1652 accuracy than generic features despite its shorter length.Table 3 reports Recall@K and average precision for the comparison.
- 5.2 Geo-localization Results: Drone-view queries outperform ground-view queries when retrieving relevant satellite-view images.The paper attributes this setting to the similar aerial viewpoints and fewer obstacles in drone-view imagery.
- 5.2 Geo-localization Results: Increasing drone-view training images from 1 to 54 per class raises both Rank@1 and AP accuracy.Figure 4 compares the Drone →Satellite and Satellite →Drone tasks across five training-set sizes.
- 5.3 Ablation Study and Further Discussion: Instance loss outperforms triplet and contrastive losses on both Drone →Satellite and Satellite →Drone tasks.The comparison uses the same backbone and satellite-view and drone-view training data, with hard sampling for the competing losses.
- 5.3 Ablation Study and Further Discussion: Sharing CNN weights improves both tasks, while separate branches are prone to overfitting the single satellite image per location.The authors state that drone-view images can regularize the shared model.
- 5.3 Ablation Study and Further Discussion: Input size 384 improves both tasks, whereas size 512 reduces performance relative to the smaller settings.The paper speculates that 512 differs too much from the 224 × 224 ImageNet pretraining size.
- 5.3 Ablation Study and Further Discussion: Model-III, which uses all annotated data, achieves the best performance in three of four tasks.Model-II performs better for drone navigation, where it is trained only on drone-view and satellite-view data.
6 CONCLUSION
The paper introduces University-1652 as a multi-view, multi-source benchmark with satellite, drone, and ground-camera data. It frames two new tasks as image retrieval and presents a baseline for learning viewpoint-invariant features.
- 6 CONCLUSION: University-1652 contains satellite, drone, and ground-camera data and enables drone-view target localization and drone navigation.The benchmark is designed for cross-view geo-localization across three platforms.
- 6 CONCLUSION: The two tasks are formulated as image retrieval problems using a baseline model that learns viewpoint-invariant features.The conclusion reports competitive performance relative to generic features and feasibility for both tasks.
A MORE QUANTITATIVE RESULTS
Additional experiments examine the effect of Google Image ground-view data on benchmark performance. Adding this noisy data generally improves both ground-related and drone-related tasks.
- A MORE QUANTITATIVE RESULTS: Adding Google Image training data generally improves Ground →Satellite, Satellite →Ground, Drone →Satellite, and Satellite →Drone performance.The authors report that most retrieved images are true matches despite noise and that the baseline handles the noisy training data.
B MORE QUALITATIVE RESULTS
A t-SNE visualization of cross-view features shows location-consistent clustering in the learned feature space. Same-location features are close, while features from different buildings are separated.
- B MORE QUALITATIVE RESULTS: Features from the same location are close and features from different target buildings are far apart in the t-SNE projection.The visualization uses 500 drone-view and satellite-view test pairs to inspect the learned feature space.
C MORE DETAILS OF UNIVERSITY-1652
University-1652 records building-level metadata, presents sample building names, and evaluates robustness to noisy Google Image training data.
- The dataset release follows Google fair-usage guidance and policies similar to existing datasets using Google data.
- The baseline trained with noisy Google Image data performs generally better across all four tasks.The ablation also reports that the baseline performs well against dataset noise.
- Each frame records the building name, longitude, latitude, altitude, heading, tilt, and range.
- Table 12 lists the first 100 building names from University-1652.