VME: A Satellite Imagery Dataset and Benchmark for Detecting Vehicles in the Middle East and Beyond

Ferda Ofli; Ingmar Weber; Noora Al-Emadi; Yin Yang

REVIEW 3 major objections 4 minor 46 references

VME: A Satellite Imagery Dataset and Benchmark for Detecting Vehicles in the Middle East and Beyond

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding a Middle East-focused training set of 4,282 satellite tiles with 113,737 labeled vehicles raises car detection accuracy on Middle Eastern imagery by about 56 percent, and consolidating it with five public datasets improves global…

desk verdict VME is a genuinely useful Middle East vehicle-detection dataset and CDSI is a solid consolidation effort, but the tile-level random split leaks scene context and makes the headline improvement numbers optimistic until re-validated. read the letter →

arxiv 2505.22353 v1 pith:6O2P4AWQ submitted 2025-05-28 cs.CV cs.AI

Noora Al-Emadi , Ingmar Weber , Yin Yang , Ferda Ofli This is my paper

classification cs.CVcs.AI

keywords satelliteimageryvehicledetectiongeographicbiasobjectdatasetorientedboundingboxestinyobjectsbenchmarkMiddleEast

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that existing satellite-imagery vehicle detectors perform poorly on Middle Eastern images because of geographic bias in their training data, not because the detectors are weak. To make this case, it introduces the Vehicles in the Middle East (VME) dataset: 4,282 high-resolution tiles from 54 cities in 12 countries, with 113,737 boxes for cars, buses, and trucks, annotated in both oriented and horizontal form. It then builds the Car Detection in Satellite Imagery (CDSI) benchmark by merging VME with five public datasets, and reports that models trained on the existing datasets alone reach only 5–56% mAP50 on VME images, while models trained on VME reach 81–86% mAP50, a relative improvement of about 56%. If true, a few thousand region-specific labeled tiles are enough to close most of the cross-region performance gap.

What carries the argument

The load-bearing objects are the two datasets and the pipeline that builds them. VME is created from 2,714 large satellite scenes, cropped into 512×512 tiles, manually filtered to keep 4,282 tiles that visibly contain objects, and annotated with oriented and horizontal boxes for car, bus, and truck. CDSI is assembled by converting five public datasets to MS-COCO horizontal-box format, mapping every car-related class to a single 'car' class when the horizontal-box area is below 400 pixels, relabeling other small objects as 'other', and pooling all images into shared train, validation, and test splits. The 400-pixel filter is what makes the cross-dataset comparison possible, because it captures more than 90% of car-related instances in every source dataset.

What would settle it

A scene-level held-out evaluation — putting all tiles from the same original source image into a single split — would settle whether the improvement over existing datasets is real; if detection accuracy on such a split falls back toward the cross-dataset baselines, the headline gain is overstated.

Watch

Extended reading notes

Core claim

The central claim is that geographic representativeness of training data is the binding constraint for satellite vehicle detection in the Middle East. Using two modern detectors, the paper shows that models trained on xView, DOTA-v2.0, VEDAI, DIOR, or FAIR1M-2.0 transfer poorly to VME images, scoring as low as 4.2% mAP50, whereas a model trained on VME itself scores 86.5% mAP50 on the same car-only test set. The paper further shows that consolidating all sources into CDSI, with car-related classes unified under a 400-pixel area filter, raises global car-detection mAP50 by 19.6% to 84.6% over models trained on any single dataset. The datasets and construction scripts are released for reuse.

Load-bearing premise

The central claim rests on the assumption that the VME test set fairly measures Middle East vehicle detection; because test tiles are random crops drawn from only 2,714 large source scenes, manually filtered to keep only tiles with obvious objects, and split at the tile level, tiles from the same source scene can appear in both training and testing, which could inflate the measured improvement.

Editorial extensions

If this is right

A detector trained only on VME outperforms every detector trained on existing public datasets when tested on Middle Eastern images, making region-specific data the decisive factor.
Combining VME with the five public datasets raises global car detection accuracy on a shared test set by tens of percentage points, so a unified benchmark is a better foundation for future work.
The CDSI conversion protocol (unified classes, a 400-pixel area filter, and MS-COCO format) gives the community a recipe for folding new regional datasets into one benchmark.
Because all car instances in VME fall in the tiny-object range below 512 pixels of area, the dataset is particularly suited to studying small-object detection.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same pattern likely holds for other under-represented regions; if so, a comparably sized regional dataset could close most of the gap there too.
Because the 512×512 test tiles are random crops from only 2,714 large scenes, tiles from the same scene can appear in both training and testing, so re-splitting at the scene level could lower the reported VME gains and would be a sharper test of the claim.
The 400-pixel car filter is calibrated to the 30–50 cm resolution of the source imagery; applying CDSI to finer imagery would require recalibrating the threshold, since real cars would occupy more pixels.
An ablation that removes the 'other' hard-negative class could reveal how much of the measured improvement comes from regional visual context rather than from extra non-car supervision.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The VME dataset is a genuinely useful new resource, and the CDSI consolidation is a real service. The paper deserves a serious referee, but the headline improvement numbers should not be taken at face value yet.

What is new: VME provides 4,282 high-resolution tiles from 54 cities in 12 Middle Eastern countries, with 113,737 vehicle instances, annotated with both OBB and HBB. The annotation QC is strong — F1 0.984 on a 5% in-house re-labeling sample. The CDSI benchmark unifies six datasets under a common car/other taxonomy and is the largest consolidated car-detection set I know of. The cross-region gap is large and visually convincing: models trained on VEDAI or FAIR1M-2.0 get single-digit mAP50 on Middle Eastern tiles while VME-trained models get above 80. That difference is big enough to survive a lot of methodological noise.

The soft spots are real, though. The main one is the test split. The dataset was built by randomly cropping 512x512 tiles from 2,714 large Maxar snapshots, then splitting the tiles randomly. Crops from the same source scene can appear in both training and test, which lets a model memorize scene-level context (road layout, parking lot structure) rather than generalize to genuinely new locations. The paper reports no source-scene-level split or overlap analysis, so the reported VME gains and the CDSI global gains are likely optimistic. The manual pre-filtering that threw out empty tiles also means the test set is not a random sample of Middle Eastern geography. The 400-pixel threshold that defines a car in CDSI anchors the benchmark to that definition, so the benchmark cannot detect whether larger cars are missed. Finally, there are no variance estimates: a single random split, no repeated seeds, so we don't know how stable the numbers are. The '56.3% improvement' in the abstract is not clearly tied to a specific table entry, which makes it hard to audit.

That said, this is a dataset paper, not a theory paper. The resource itself is solid, the documentation is complete, and the data is released. The qualitative conclusion — geographic bias, not architecture, drives the cross-region gap — is likely correct. The quantitative claims need re-validation. I'd send it to review with a request for scene-level splits, confidence intervals, and a clearer derivation of the headline number. The target audience is remote sensing and geospatial CV researchers; they will use this dataset regardless. I'd bring it to reading group — the split-leakage discussion is a good teaching moment.

Referee Report

3 major / 4 minor

Summary. The paper presents VME, a new satellite-imagery dataset of 4,282 512x512 tiles from 54 Middle Eastern cities (Maxar, 30-50 cm GSD) with 113,737 OBB/HBB vehicle annotations (car/bus/truck), and CDSI, a consolidated benchmark of 23,250 images that merges VME with car-related classes from xView, DOTA-v2.0, VEDAI, DIOR, and FAIR1M-2.0 after filtering to a two-class car/other taxonomy. The authors report high annotation quality (F1 0.984 on a 5% sample), baselines with TOOD and DINO Swin-L under SAHI, and results showing models trained on existing datasets transfer poorly to VME (e.g., 5.1-53.3 mAP50 car), while VME-trained models reach 81.2-86.5 mAP50 and CDSI-trained models improve global car detection by 19.6-84.6% relative to individual datasets. The central claim is that geographic bias, rather than detector architecture, causes the cross-region gap, and that adding Middle Eastern training data mostly closes it.

Significance. The dataset addresses a genuine gap: Middle Eastern cities are strongly underrepresented in current public benchmarks, and the release includes both OBB and HBB annotations, source code, and construction scripts. The annotation QC procedure is a strength, and the consolidation of five existing datasets into a unified car benchmark is a useful resource. However, the headline quantitative claims are only as strong as the evaluation splits; the current tile-level random split does not establish cross-scene generalization, so the benchmark numbers should be treated as upper bounds until a scene-disjoint evaluation is provided.

major comments (3)

[Methods, 'Final Dataset' and 'Detection Benchmarks'] The VME train/val/test split is made by randomly dividing the 4,282 tiles after they were cropped from 2,714 large Maxar snapshots. A single snapshot contributes many tiles (22,125 tiles from 2,714 snapshots before filtering), so the same geographic scene can appear in both training and test. This violates the independence that a held-out test set should provide; the reported VME gains (Table 5, e.g., TOOD car: 81.2 vs 53.3, DINO: 86.5 vs 62.3) may be inflated by scene-memorization. Please quantify the number of source snapshots that straddle the split, report results on a snapshot-disjoint split, and ideally a held-out city split; if the improvement over existing datasets persists, the central claim will be substantially strengthened.
[Methods, 'CDSI Benchmark' and Table 5] The CDSI test set is the union of the test splits of all constituent datasets, including the VME test tiles, and the CDSI training set includes the VME training tiles. Therefore the comparison between CDSI and CDSI* on the CDSI test set does not measure generalization to unseen global scenes; it measures in-domain performance on a test partition that shares source scenes with training. Please add an evaluation on a held-out set of scenes from outside the constituent datasets (e.g., unseen cities/countries from Maxar or an external benchmark) and report both CDSI and CDSI* there. This is necessary to support the 'global car detection' claim.
[Methods, 'Image Collection' and 'Final Dataset'] The manual discarding of tiles with 'no obvious objects' (completely green or desert areas) removes exactly the empty background that a detector must reject in operational use. Because the test set only contains tiles that passed this filter, absolute mAP values are not representative of full-scene performance. Please report the number and proportion of discarded tiles per city, and provide at least one whole-snapshot or unfiltered-tile evaluation to bound the optimistic bias.

minor comments (4)

[Tables 1 and 3] The VME test split contains 1,011 images in Table 1 but 988 images in Table 3, and the VME original-set totals in Table 3 sum to 4,018 rather than the 4,282 images described in the text. Please reconcile these counts and clarify the treatment of the 241 no_label images.
[Table 5, VEDAI car-other row] For the TOOD model on the VME test set, the table reports mAP50=4.0 with mAP=9.0, which is impossible under standard COCO evaluation; please correct the entry or explain the protocol.
[Background & Summary] The text claims a 'remarkable 56.3% improvement' for VME car detection, but Table 5 does not directly reproduce this number from any pair of rows. Please state the exact comparison used (which setup, which metric, which baseline).
[Methods, 'Category Mapping'] The 400-pixel HBB-area filter is used both to define the car class in CDSI and to filter VME's own car annotations. Because all benchmark numbers in Table 5 are for this filtered class, the paper should state explicitly that the benchmark does not evaluate detection of larger vehicles that some source taxonomies label as cars.

Circularity Check

1 steps flagged · score 2.0 of 10

No circular derivation; the only mildly self-definitional element is that CDSI's 'car' class is defined by the same 400-pixel area filter that the CDSI benchmark then measures, so larger vehicles are excluded by construction.

self definitional [Methods: CDSI Dataset, 'Data Processing and Filtering' and 'Category Mapping'; Technical Validation, 'CDSI Benchmark'.]
"Given that we are interested in a GSD range of 30-50cm per pixel, we assume that an object with an area greater than 400 pixels is unlikely a car. ... This analysis reveals that an area size of less than 400 pixels reports for more than 90% of all car-related object instances across all datasets. Therefore, we decided to filter out all object instances with an area larger than 400 pixels; even if they were originally labeled as cars."

The CDSI benchmark then uses CDSI test images built with the same filter to support the claim that CDSI yields 'substantial improvements in global car detection.' Since every object in the car class of CDSI is, by construction, an object with HBB area below 400 pixels, the benchmark can only measure detection of that threshold-defined class. Instances that other datasets originally labeled as cars but that are larger than 400 pixels (e.g., vans, large SUVs) are excluded from both training labels and test ground truth, so the headline improvement cannot detect or correct any failure on them.

full rationale

This is a dataset/benchmark paper; its central claims are empirical mAP numbers, not a mathematical derivation. I checked for the standard circularity patterns. No fitted parameter is relabeled as a prediction: all models are trained and then evaluated on fixed test splits. No load-bearing self-citation: reference 39 is the dataset's own Zenodo record and reference 10 is motivational prior work; neither is used to force the benchmark conclusions. No uniqueness theorem or ansatz is imported from the authors' prior work. The one self-definitional anchor is the 400-pixel threshold used to build CDSI's 'car' class and then to measure CDSI's car-detection performance; that limits the scope of the claim but does not by itself manufacture the reported accuracy gains. The more serious threat to the headline numbers is not circular but statistical: the VME test set is randomly cropped from the same 2,714 Maxar snapshots as the training tiles, so same-scene context can leak across splits; this is a data-leakage/correctness concern and should be addressed with scene-level splits, but it is not an equivalence-to-inputs circularity. Overall score 2.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central results rest on annotation quality, the cross-dataset class mapping threshold, and the validity of the VME test split. No new physical or theoretical entities are introduced; VME and CDSI are datasets, not theoretical constructs.

free parameters (1)

CDSI car size threshold (HBB area < 400 pixels) = 400 px
Used in Sec. 2.3 to convert car-related classes from five datasets into a single car class and to relabel other small objects. It is justified by a histogram analysis but is a hand-chosen cutoff that defines the CDSI label space and all benchmark evaluations.

assumptions (5)

domain assumption High-resolution satellite images at 30 to 50 cm/pixel with <20% cloud cover and off-nadir angle <30 degrees allow reliable per-vehicle annotation.
This underpins the entire VME collection and the claim that more than 100,000 objects are correctly labeled (Sec. 2.1 Image Collection).
domain assumption Crowdsourced annotations from Co-one, after expert review and validation on a 5% sample, are accurate enough to serve as ground truth for benchmarking.
The quality check reports precision 0.999 and recall 0.970, but only 215 images are spot-checked and the reference labels come from a single annotator (Sec. 3.1).
ad hoc to paper Car-related classes across xView (small car), DOTA-v2.0 (small vehicle), VEDAI (car, van), DIOR (vehicle), and FAIR1M-2.0 (small car, van) denote the same physical object after applying the 400-pixel area filter.
This mapping is specific to this paper's CDSI design and is not validated by an external benchmark (Sec. 2.2 Category Mapping and Sec. 2.3 Data Processing).
domain assumption The label distributions of the five existing datasets are representative of global car imagery for comparison purposes.
The geographic-bias claim compares Middle East performance with models trained on these datasets, but their geographic coverage is uneven and mostly unreported (Sec. 2.2 and Fig. 2).
ad hoc to paper Random splits at the tile level, not the source-scene level, provide independent training and test sets.
The paper creates 512x512 crops from 2,714 large images and then splits the crops randomly. This assumption underlies the reliability of the VME and CDSI test-set numbers and is likely violated when crops from the same scene appear in both splits (Sec. 2.1 and Final Dataset).

how reviews work

0 comments

Cite this review

Pith. "Pith review of VME: A Satellite Imagery Dataset and Benchmark for Detecting Vehicles in the Middle East and Beyond." pith.science (2026). https://pith.science/paper/6O2P4AWQ

@misc{pith2026250522353,
  author       = {Pith},
  title        = {Pith review of: VME: A Satellite Imagery Dataset and Benchmark for Detecting Vehicles in the Middle East and Beyond},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6O2P4AWQ}},
  note         = {Machine review of arXiv:2505.22353}
}

read the original abstract

Detecting vehicles in satellite images is crucial for traffic management, urban planning, and disaster response. However, current models struggle with real-world diversity, particularly across different regions. This challenge is amplified by geographic bias in existing datasets, which often focus on specific areas and overlook regions like the Middle East. To address this gap, we present the Vehicles in the Middle East (VME) dataset, designed explicitly for vehicle detection in high-resolution satellite images from Middle Eastern countries. Sourced from Maxar, the VME dataset spans 54 cities across 12 countries, comprising over 4,000 image tiles and more than 100,000 vehicles, annotated using both manual and semi-automated methods. Additionally, we introduce the largest benchmark dataset for Car Detection in Satellite Imagery (CDSI), combining images from multiple sources to enhance global car detection. Our experiments demonstrate that models trained on existing datasets perform poorly on Middle Eastern images, while the VME dataset significantly improves detection accuracy in this region. Moreover, state-of-the-art models trained on CDSI achieve substantial improvements in global car detection.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 37 canonical work pages

[1]

Nguyen, T. T. et al. Monitoring agriculture areas with satellite images and deep learning. Applied Soft Computing 95, 106565 (2020)

work page 2020
[2]

Wang, Y ., Cai, G., Y ang, L., Zhang, N. & Du, M. Monitoring of urban ecological environment including air quality using satellite imagery. Plos one 17, e0266759 (2022)

work page 2022
[3]

& Gonzalez, M

Albert, A., Kaur, J. & Gonzalez, M. C. Using convolutional networks and satellite imagery to identify patterns in urban environments at a large scale. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 1357–1366 (2017)

work page 2017
[4]

Huang, X. et al. Urban building classification (ubc)-a dataset for individual building detection and classification from satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1413–1421 (2022)

work page 2022
[5]

Toward more integrated utilizations of geostationary satellite data for disaster management and risk mitigation

Higuchi, A. Toward more integrated utilizations of geostationary satellite data for disaster management and risk mitigation. Remote Sensing 13, 1553 (2021)

work page 2021
[7]

& de Franchis, C

Drouyer, S. & de Franchis, C. Highway traffic monitoring on medium resolution satellite images. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, 1228–1231 (IEEE, 2019)

work page 2019
[8]

& Albanwan, H

Chen, Y ., Qin, R., Zhang, G. & Albanwan, H. Spatial temporal analysis of traffic patterns during the covid-19 epidemic by vehicle detection using planet remote-sensing satellite images. Remote Sensing 13, 208 (2021)

work page 2021
[9]

& Orlikova, L

Golej, P ., Horak, J., Kukuliac, P . & Orlikova, L. Vehicle detection using panchromatic high-resolution satellite images as a support for urban planning. case study of prague’s centre.GeoScape 16 (2022)

work page 2022

Show all 46 references

[10]

& Weber, I

Rufener, M.-C., Ofli, F ., Fatehkia, M. & Weber, I. Estimation of internal displacement in ukraine from satellite-based car detections. Sci. Reports 14, 31638 (2024)

2024
[11]

Liu, H.-I. et al. A denoising fpn with transformer r-cnn for tiny object detection. IEEE Transactions on Geoscience and Remote Sensing (2024)

2024
[12]

Verma, T. et al. Soar: Advancements in small body object detection for aerial imagery using state space models and programmable gradients. Preprint at https://doi.org/10.48550/arXiv.2405.01699 (2024)

2024 doi arXiv
[13]

Zhu, J. et al. Transformer based remote sensing object detection with enhanced multispectral feature extraction. IEEE Geoscience and Remote Sensing Letters (2023)

2023
[14]

& Zhou, F

Tong, K., Wu, Y . & Zhou, F . Recent advances in small object detection based on deep learning: A review. Image and Vision Computing 97, 103910 (2020)

2020
[15]

Cheng, G. et al. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

2023
[16]

& Tian, J

Gao, P ., Tian, T., Li, L., Ma, J. & Tian, J. De-cyclegan: An object enhancement network for weak vehicle detection in satellite images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 3403–3414 (2021)

2021
[17]

& Li, H.-C

Du, Q., Celik, T., Wang, Q. & Li, H.-C. Fully convolutional lightweight pyramid network for vehicle detection in aerial images. IEEE Geoscience and Remote Sensing Letters (2021)

2021
[18]

Li, X. et al. Vehicle detection in very-high-resolution remote sensing images based on an anchor-free detection model with a more precise foveal area. ISPRS International Journal of Geo-Information 10, 549 (2021)

2021
[19]

& Zhang, T

Shi, F ., Zhang, T. & Zhang, T. Orientation-aware vehicle detection in aerial images via an anchor-free object detection approach. IEEE Transactions on Geoscience and Remote Sensing 59, 5221–5233 (2020)

2020
[20]

Ding, J. et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence 44, 7778–7796, https://doi.org/10.1109/TPAMI.2021.3117983 (2021)

2021
[21]

& Jurie, F

Razakarivony, S. & Jurie, F . Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation 34, 187–203, https://doi.org/10.1016/j.jvcir.2015.11.002 (2016)

2016 doi
[22]

Lam, D. et al. xView: Objects in context in overhead imagery. Preprint at https://doi.org/10.48550/arXiv.1802.07856 (2018)

2018 doi arXiv
[23]

& Mukherjee, R

Christie, G., Fendley, N., Wilson, J. & Mukherjee, R. Functional map of the world. In CVPR (2018)

2018
[24]

& Li, C.-Y

Lin, H.-Y ., Tu, K.-C. & Li, C.-Y . Vaid: An aerial image dataset for vehicle detection and classification. IEEE Access 8, 212209–212219 (2020)

2020
[25]

& Xia, G.-S

Wang, J., Y ang, W ., Guo, H., Zhang, R. & Xia, G.-S. Tiny object detection in aerial images. In 2020 25th international conference on pattern recognition (ICPR), 3791–3798 (IEEE, 2021)

2020
[26]

P ., Rotich, G

Minetto, R., Segundo, M. P ., Rotich, G. & Sarkar, S. Measuring human and economic activity from satellite imagery to support city- scale decision-making during covid-19 pandemic. IEEE Transactions on Big Data 7, 56–68 (2020)

2020
[27]

7 Impressive Smart City Projects in the Middle East

ZIGURAT Institute of Technology, Z. 7 Impressive Smart City Projects in the Middle East. https://www.e-zigurat.com/en/blog/ smart-city-projects-middle-east/ Accessed on 2024-09-17 (2023)

2023
[28]

The Rise of Gulf Smart Cities

George, R. The Rise of Gulf Smart Cities. Wilson Center. Accessed on 2024-09-18 https://www.wilsoncenter.org/article/rise-gulf- smart-cities Accessed on 2024-09-18 (2024)

2024
[29]

& Han, J

Li, K., Wan, G., Cheng, G., Meng, L. & Han, J. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing 159, 296–307, https://doi.org/10.1016/j.isprsjprs.2019.11.023 (2020)

2020 doi
[30]

Sun, X. et al. FAIR1M: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 184, 116–130, https://doi.org/10.1016/j.isprsjprs.2021.12.004 (2022)

2022 doi
[31]

Feng, C., Zhong, Y ., Gao, Y ., Scott, M. R. & Huang, W . Tood: Task-aligned one-stage object detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 3490–3499 (IEEE Computer Society, 2021)

2021
[32]

Zhang, H. et al. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations (ICLR) (2023)

2023
[33]

N., Konjevod, G., Sakla, W

Mundhenk, T. N., Konjevod, G., Sakla, W . A. & Boakye, K. A large contextual dataset for classification, detection and counting of cars with deep learning. In European Conference on Computer Vision, 785–800 (Springer, 2016)

2016
[34]

Zambanini, S., Loghin, A.-M., Pfeifer, N., Soley, E. M. & Sablatnig, R. Detection of parking cars in stereo satellite images. Remote Sensing 12, 2170 (2020)

2020
[35]

& Benjdira, B

Ammar, A., Koubaa, A., Ahmed, M., Saad, A. & Benjdira, B. Vehicle detection from aerial images using deep learning: A comparative study. Electronics 10, 820 (2021)

2021
[36]

Zhu, P . et al. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 7380–7399 (2021)

2021
[37]

VehSat: a large-scale dataset for vehicle detection in satellite images

Drouyer, S. VehSat: a large-scale dataset for vehicle detection in satellite images. In IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, 268–271, https://doi.org/10.1109/IGARSS39084.2020.9323289 ISSN: 2153-7003 (2020)

2020
[38]

M., Bahmanyar, R., Henry, C

Azimi, S. M., Bahmanyar, R., Henry, C. & Kurz, F . Eagle: Large-scale vehicle detection dataset in real-world scenarios using aerial imagery. In 2020 25th International Conference on Pattern Recognition (ICPR), 6920–6927 (IEEE, 2021)

2020
[39]

& Ofli, F

Al-Emadi, N., Weber, I., Y ang, Y . & Ofli, F . VME: A Satellite Imagery Dataset and Benchmark for Detecting Vehicles in the Middle East and Beyond. https://doi.org/10.5281/zenodo.14185684 (2024)

2024 doi
[40]

C., Altinuc, S

Akyon, F . C., Altinuc, S. O. & Temizel, A. Slicing aided hyper inference and fine-tuning for small object detection. In 2022 IEEE International Conference on Image Processing (ICIP), 966–970 (IEEE, 2022)

2022
[41]

Tian, Z., Shen, C., Chen, H. & He, T. FCOS: Fully Convolutional One-Stage Object Detection. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 9626–9635, https://doi.org/10.1109/ICCV .2019.00972 (IEEE, 2019)

2019
[42]

& Sunderhauf, N

Zhang, H., Wang, Y ., Dayoub, F . & Sunderhauf, N. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8514–8523 (2021)

2021
[43]

Chen, K. et al . MMDetection: Open mmlab detection toolbox and benchmark. Preprint at https://doi.org/10.48550/ arXiv.1906.07155 (2019)

2019 doi arXiv
[44]

Jakubik, J. et al . Foundation models for generalist geospatial artificial intelligence. Preprint at https://doi.org/10.48550/ arXiv.2310.18660 (2023)

2023 doi arXiv
[45]

Jakubik, J. et al. HLS Foundation, Prithvi-100M, https://doi.org/10.57967/hf/0952 (2023)

2023 doi
[46]

& Hinton, G

Van der Maaten, L. & Hinton, G. Visualizing data using t-sne. Journal of machine learning research 9 (2008)

2008
[47]

& Dai, Q

Hoiem, D., Chodpathumwan, Y . & Dai, Q. Diagnosing error in object detectors. In European conference on computer vision, 340–353 (Springer, 2012). 14Scientific Data | (2025) 12:500 | https://doi.org/10.1038/s41597-025-04567-y www.nature.com/scientificdatawww.nature.com/scienti...

2025 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Nguyen, T. T. et al. Monitoring agriculture areas with satellite images and deep learning. Applied Soft Computing 95, 106565 (2020)

work page 2020

[2] [2]

Wang, Y ., Cai, G., Y ang, L., Zhang, N. & Du, M. Monitoring of urban ecological environment including air quality using satellite imagery. Plos one 17, e0266759 (2022)

work page 2022

[3] [3]

& Gonzalez, M

Albert, A., Kaur, J. & Gonzalez, M. C. Using convolutional networks and satellite imagery to identify patterns in urban environments at a large scale. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 1357–1366 (2017)

work page 2017

[4] [4]

Huang, X. et al. Urban building classification (ubc)-a dataset for individual building detection and classification from satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1413–1421 (2022)

work page 2022

[5] [5]

Toward more integrated utilizations of geostationary satellite data for disaster management and risk mitigation

Higuchi, A. Toward more integrated utilizations of geostationary satellite data for disaster management and risk mitigation. Remote Sensing 13, 1553 (2021)

work page 2021

[6] [7]

& de Franchis, C

Drouyer, S. & de Franchis, C. Highway traffic monitoring on medium resolution satellite images. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, 1228–1231 (IEEE, 2019)

work page 2019

[7] [8]

& Albanwan, H

Chen, Y ., Qin, R., Zhang, G. & Albanwan, H. Spatial temporal analysis of traffic patterns during the covid-19 epidemic by vehicle detection using planet remote-sensing satellite images. Remote Sensing 13, 208 (2021)

work page 2021

[8] [9]

& Orlikova, L

Golej, P ., Horak, J., Kukuliac, P . & Orlikova, L. Vehicle detection using panchromatic high-resolution satellite images as a support for urban planning. case study of prague’s centre.GeoScape 16 (2022)

work page 2022