University-1652: Drone-Based Cross-View Benchmark

• University-1652 is a multi-view, multi-source benchmark dataset capturing 1,652 university buildings with images from satellite, drone, and ground platforms.
• It supports retrieval tasks like drone-view target localization and drone navigation using instance loss and evaluation metrics such as Recall@K and Average Precision.
• The dataset has spurred advanced methods including diffusion-based localization, weather-robust map fusion, and multimodal language-guided spatial reasoning.

University-1652 is a multi-view, multi-source benchmark for drone-based cross-view geo-localization introduced in "University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization" (Zheng et al., 2020). It contains imagery of 1,652 university buildings from 72 universities worldwide, with each building treated as a semantic class and observed from satellite, drone, and ground platforms. The dataset was constructed to support learning of viewpoint-invariant features under drastic viewpoint change, especially for two drone-centric retrieval tasks: drone-view target localization and drone navigation. Subsequent work has used University-1652 not only as a benchmark for image-to-image retrieval, but also as a substrate for diffusion-based localization, weather-robust map augmentation, and language-guided spatial reasoning.

1. Dataset definition and visual composition

University-1652 is explicitly characterized as both multi-view and multi-source. “Multi-source” refers to three imaging platforms: satellites, drones, and ground/phone cameras. “Multi-view” refers to multiple distinct viewpoints around the same building, rather than a single paired observation. This design contrasts with earlier geo-localization datasets that typically provide only one ground image and one aerial image per location. The central argument of the original benchmark is that drones provide a third platform that is less occluded than ground imagery, more similar in scale and viewpoint to satellite imagery, and controllable enough to trace a smooth viewpoint transition between street-level and overhead observations (Zheng et al., 2020).

Platform	Images per building	Source and acquisition
Satellite	1	Top-down image from geo-coordinates
Drone	54	Google Earth 3D engine; spiral flight with 3 rounds, descending from 256 m to 121.5 m; one frame sampled every 15 frames from 30 fps video
Ground	3.38 street-view on average; 16.64 ground images per building in training if including Google Image	Google Street View plus Google Image search, with indoor-image filtering and duplicate removal

The drone-view stream is synthetic rather than physically captured at scale. The benchmark uses the Google Earth 3D engine to generate a spiral trajectory around each building, with three loops and descending altitude. The ground-view portion combines Google Street View with extra web images queried by building name. The added ground images are filtered by a ResNet-18 classifier trained on Places to remove indoor images, and duplicates are removed following a prior duplicate-removal procedure. The dataset deliberately focuses on university campus buildings rather than landmarks, both to reduce bias toward highly distinctive architecture and because drones are often forbidden near landmarks (Zheng et al., 2020).

Each image is associated with a building ID, and drone frames additionally record frame-level metadata including building name, longitude, latitude, altitude, heading, tilt, and range. Those metadata are not used in the original experiments, but they were presented as available for future work on orientation-aware or view-aware modeling (Zheng et al., 2020).

2. Split design, retrieval tasks, and evaluation protocol

The original construction distinguishes between buildings with full three-view coverage and partial-view distractors. There are 1,402 buildings that have satellite, drone, and ground views; these are evenly split into 701 training buildings from 33 universities and 701 test buildings from 39 universities. The remaining 250 buildings lack complete three-view coverage and are used as gallery distractors. A key property of the split is that universities do not overlap between train and test, so evaluation measures cross-campus generalization rather than memorization of campus-specific appearance (Zheng et al., 2020).

The detailed split reported in the benchmark consists of 50,218 training images over 701 classes; a test query set of 37,855 drone images, 701 satellite images, and 2,579 ground images; and gallery sets of 51,355 drone images over 951 classes, 951 satellite images over 951 classes, and 2,921 ground images over 793 classes. In the original interpretation, the 951 gallery buildings are the 701 test buildings plus 250 distractors. Later papers sometimes restate the benchmark more compactly as 701 training buildings and 951 testing buildings, reflecting the gallery-oriented test construction rather than the original full/partial-view distinction.

University-1652 enables two primary retrieval tasks. Drone-view target localization is the drone-to-satellite setting: a drone image is the query and the satellite gallery is searched for the matching building. Drone navigation is the satellite-to-drone setting: a satellite image is the query and the drone gallery is searched for views of the same building. The benchmark also supports ground-to-satellite, satellite-to-ground, ground-to-drone, and drone-to-ground retrieval, but the drone–satellite pair is the defining use case. In these tasks, positives are gallery items with the same building ID, while negatives are images from other buildings, including distractors. Because the dataset is multi-view, a query may have multiple correct matches in the gallery.

Evaluation uses both Recall@K and Average Precision. Recall@K is defined in the standard retrieval sense,

$$\mathrm{Recall@K}=\frac{1}{N_q}\sum_{i=1}^{N_q} R_i@K,$$

where $R_i@K$ indicates whether at least one true match appears within the top $K$ ranked gallery items for query $i$. The benchmark also reports AP and mean AP because multiple positives per query make single-hit recall incomplete as a summary of ranking quality. A multiple-query setting is additionally defined for drone queries: features from multiple drone images of the same building are averaged by simple mean pooling before retrieval (Zheng et al., 2020).

3. Baseline modeling and early empirical findings

The original strong baseline is a ResNet-50 pretrained on ImageNet, modified by removing the original classifier, adding a 512-dimensional fully connected embedding layer, and attaching a shared classification layer over building IDs. For multi-source training, the authors define separate branches for satellite, drone, and ground views, but in the full three-branch model the satellite and drone branches share weights while the ground branch remains distinct. This weight sharing is motivated by the fact that both satellite and drone views are aerial and that the single satellite image per building makes a fully separate satellite branch prone to overfitting (Zheng et al., 2020).

The key loss is instance loss, a classification-style objective in which each building is a class and the same shared classifier weights are used across views. The shared classifier imposes a soft alignment constraint across source-specific feature extractors without explicitly enforcing triplet or contrastive metric constraints. In a controlled comparison using only satellite and drone data, instance loss outperformed contrastive loss, triplet loss with margins 0.3 and 0.5, and weighted soft-margin triplet loss. For Drone $\rightarrow$ Satellite, instance loss achieved $R@1=58.23$ and $AP=62.91$; for Satellite $\rightarrow$ Drone, it achieved $R@1=74.47$ and $AP=59.45$.

The benchmark also established that training directly on University-1652 is materially more effective than using generic off-the-shelf features. ResNet-based features trained on ImageNet, Places365, or SfM-120k produced Drone $R_i@K$0 Satellite $R_i@K$1 values around 5–13% and AP values around 7–16%, whereas a feature learned on University-1652 reached $R_i@K$2 and $R_i@K$3 for Drone $R_i@K$4 Satellite and $R_i@K$5 and $R_i@K$6 for Satellite $R_i@K$7 Drone. Multiple-query drone retrieval further improved performance to $R_i@K$8 and $R_i@K$9, while Ground $K$0 Satellite remained very weak at $K$1 and $K$2 in the single-query setting. The authors also reported that increasing the number of drone training images per class from $K$3 steadily improved both retrieval directions, and they showed qualitative generalization from synthetic training data to a real 4K drone video from an unseen test-set school, as well as transfer gains on Oxford, Paris, Revisited Oxford, and Revisited Paris place-recognition benchmarks (Zheng et al., 2020).

4. Diffusion-based cross-view localization on University-1652

"DiffusionUavLoc: Visually Prompted Diffusion for Cross-View UAV Localization" treats University-1652 as the primary UAV–satellite benchmark for an image-prompted, text-free, diffusion-centric localization framework (Liu et al., 9 Nov 2025). In that work, the dataset is used both as a training corpus for supervised UAV-to-satellite synthesis and latent alignment and as an evaluation benchmark for both Drone $K$4 Satellite and Satellite $K$5 Drone retrieval. The ground/street-view images are not used. The paper emphasizes that University-1652 is the more challenging and central benchmark relative to SUES-200 because it presents large viewpoint differences between strongly oblique UAV images and near-nadir satellite tiles, along with complex occlusions and structural clutter around buildings.

The method first applies scene-wise processing to UAV imagery. A COLMAP reconstruction pipeline produces 3D point clouds, which are orthoprojectively rendered into pseudo-satellite orthophotos. The rendered orthophotos are then cleaned with a fixed 20% center crop to remove black borders or inpainting artifacts. If COLMAP fails for some scenes, a perspective-based orthorectification fallback is used. These pseudo-satellite images act as structural prompts for a text-free conditional diffusion model. A ResNet-18 encoder and projector map the pseudo-satellite prompt into the Stable Diffusion text-embedding space, while Canny edges, SAM-2 semantic masks, and DepthAnything-v2 relative depth are concatenated and encoded into a ControlNet condition. The model uses the Stable Diffusion VAE latent as a shared embedding space: a UAV image is mapped to a predicted satellite-domain latent, a real satellite image is encoded to a latent posterior mean, both descriptors are flattened and $K$6-normalized, and retrieval is performed by cosine similarity at a fixed inference timestep. Notably, the system does not train a separate retrieval network; discriminative behavior emerges from generative supervision in the satellite domain.

Training uses a multi-objective uncertainty-weighted loss composed of VGG-16 perceptual loss, stationary-wavelet detail loss, and a structural mask loss over structure-sensitive regions. On University-1652, the full model reports Drone $K$7 Satellite $K$8 and $K$9, and Satellite $i$0 Drone $i$1 and $i$2, with the latter two marked as best among compared methods. The paper also reports a macro average of 96.50 across the four directional metrics, ahead of CGSI at 95.88 and QDFL at 95.64. Ablations on University-1652 show large degradation under perspective-only projection ($i$3 for Drone $i$4 Satellite and $i$5 for Satellite $i$6 Drone), weaker performance when text prompts are reintroduced, and reduced AP when the wavelet loss is removed. The study explicitly concludes that orthographic rendering is essential for handling the dataset’s geometry, text-free visual prompting is stronger than text prompting, and wavelet detail loss materially improves ranking quality. A plausible implication is that University-1652 remains a stringent benchmark for geometry-aware cross-view alignment rather than a saturated retrieval testbed.

5. Weather robustness and road-map augmentation

"Road Maps as Free Geometric Priors: Weather-Invariant Drone Geo-Localization with GeoFuse" uses University-1652 as the principal non-dense benchmark for all-weather drone geo-localization and augments it with a new map modality (Fang et al., 14 May 2026). The core extension is a geo-aligned road-map tile paired with each satellite image. Geographic coordinates are extracted from the public benchmark, Google Maps API is queried at a zoom level chosen to balance road-network visibility and spatial coverage, and the resulting tiles are aligned geographically with the corresponding satellite images. Because raw map tiles contain street names and labels, the authors remove all textual content using Qwen-Image-Edit and retain text-free road maps as the default modality to avoid information leakage and unrealistic gains from textual cues.

GeoFuse evaluates both Drone $i$7 Satellite and Satellite $i$8 Drone retrieval under 10 synthetic weather conditions applied only to drone images: Normal, Fog, Rain, Snow, Fog+Rain, Fog+Snow, Rain+Snow, Dark, Over-exp, and Wind. Satellite imagery and road maps remain unchanged. The model fuses satellite and road-map features through token-level and channel-level interactions with dynamic gating, then aligns weather-degraded drone features to both satellite and fused satellite-road anchors using class-level cross-view contrastive learning, where each building ID is treated as a class. In the reported means across 10 weathers on University-1652, the WeatherPrompt-style X-VLM baseline achieves $i$9 and $\rightarrow$0 for Drone $\rightarrow$1 Satellite, while GeoFuse reaches $\rightarrow$2 and $\rightarrow$3, an improvement of $\rightarrow$4 in $\rightarrow$5 and $\rightarrow$6 in AP. For Satellite $\rightarrow$7 Drone, the same comparison is $\rightarrow$8 versus $\rightarrow$9, yielding gains of $R@1=58.23$0 in $R@1=58.23$1 and $R@1=58.23$2 in AP.

An auxiliary-modality ablation compares pseudo-depth maps, road maps with text, and road maps without text. Pseudo-depth is least effective; road maps clearly help; text-bearing maps slightly improve mean $R@1=58.23$3 in some settings but are not used for the main benchmark because of fairness concerns and possible semantic leakage. The paper also notes method-level limitations that bear directly on the dataset variant: if road maps are unavailable, sparse, outdated, or misaligned, the benefit of the added modality can diminish, and the advantage is expected to be greatest in built environments such as the urban campus scenes that dominate University-1652.

6. Language and spatially annotated extensions

"Towards Natural Language-Guided Drones: GeoText-1652 Benchmark with Spatial Relation Matching" turns University-1652 into a multimodal benchmark by adding text and region annotations to the original image collection (Chu et al., 2023). The extension preserves the underlying building IDs and university structure while adding one-to-one correspondences between image, text, and bounding-box elements. Every image receives global image-level descriptions, and many images additionally receive region-level descriptions paired with bounding boxes. Reported corpus statistics are 316,335 global descriptions and 276,045 bbox–text pairs, with averages of 3 global descriptions and 2.62 bounding boxes per image; global descriptions have an average length of 70.23 words, while region-level descriptions average 21.6 words.

The annotation pipeline is a human–computer loop rather than fully manual curation. A visual LLM such as MiniGPT-4 is prompted with “salient object” and “whole image” prompts to generate candidate captions. A referee model, guided by GPT-4 and implemented as positive/negative keyword filtering, removes outputs with artifacts such as URLs or placeholder phrases and enforces the presence of spatial relation tokens such as “left,” “right,” “upper,” and “down” where appropriate. Grounding DINO is then used to infer bounding boxes from region-level descriptions, and explicit spatial rules reject region–text pairs whose predicted box placement contradicts the text. The paper reports five rounds of refinement in which 20% of annotations are sampled and rated by human annotators until “excellent” annotations exceed 90%, after which the final set is pruned to high-quality matches.

GeoText-1652 defines two cross-modal retrieval tasks on top of the University-1652 image backbone: text query $R@1=58.23$4 image retrieval for language-guided drone navigation, and image query $R@1=58.23$5 text retrieval for drone-view target localization via text. The modeling framework builds on XVLM and adds two fine-grained spatial objectives. The grounding loss supervises predicted bounding boxes for region-level descriptions using IoU loss plus $R@1=58.23$6 regression, while the spatial loss classifies pairwise relations among regions into 9 categories derived from horizontal and vertical relative position. The full objective is

$R@1=58.23$7

with $R@1=58.23$8. On the test set, the full method reports $R@1=58.23$9 for text $AP=62.91$0 image retrieval and $AP=62.91$1 for image $AP=62.91$2 text retrieval, modestly improving over fine-tuned XVLM. The paper also shows that using all three original University-1652 platforms—satellite, drone, and ground—outperforms training on drone-only or satellite+ground subsets, reinforcing the original premise that multi-view, multi-source coverage is central to viewpoint-invariant learning.

The language extension also clarifies a recurring misconception about the benchmark’s scope. University-1652 was originally a visual-only image retrieval dataset, not a language benchmark, and GeoText-1652 does not replace that role; it overlays a new layer of multimodal supervision on the same visual core. At the same time, the extension inherits the original dataset’s limitations: campus-only scene coverage, structured built environments, and potential residual annotation noise from LLM-assisted captioning. Taken together with diffusion-based and weather-robust augmentations, these later works position University-1652 as a persistent experimental core for cross-view geo-localization research across image retrieval, generative modeling, multimodal fusion, and spatially grounded vision–language learning.