Pith. sign in

REVIEW 3 major objections 6 minor 58 references

Visual Place Recognition for Large-Scale UAV Applications

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A large aerial dataset and rotation-aware networks lift UAV place recognition recall by 12 percent.

desk verdict A genuinely large-scale aerial vPR dataset and a sensible steerable-CNN baseline, but the paper's central 'no overlap' design claim is unverifiable as written and the training comparison conflates scale with source diversity. read the letter →

arxiv 2507.15089 v1 pith:XMGYRF5Q submitted 2025-07-20 cs.CV cs.RO

Ioannis Tsampikos Papapetros , Ioannis Kansizoglou , Antonios Gasteratos This is my paper
classification cs.CVcs.RO
keywords visualplacerecognitionunmannedaerialvehiclesimagerylarge-scaledatasetLASEDsteerableCNNrotationinvarianceRecall@1benchmarking
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 argues that aerial visual place recognition is held back by two fixable problems: a shortage of large high-altitude training data and poor handling of image rotation. It introduces LASED, about one million orthophoto images sampled from 170,000 Estonian locations across a decade, with a structured no-overlap design, plus a Swiss test set, LASED-Test. Benchmarking shows that models trained on LASED beat the same architectures trained on existing smaller aerial datasets on every test set. The paper also plugs steerable CNNs into the pipeline, and their rotation-equivariant backbones improve average Recall@1 by 12% over the best conventional network when trained on LASED. Together the dataset and the architecture are claimed to make country-scale UAV visual localization practical.

What carries the argument

The two load-bearing pieces are LASED's sampling scheme and steerable convolutional layers. LASED samples Estonia on a constant latitude/longitude grid, filters out locations covered by water, and emits circular 500-pixel images covering roughly 400 meters of ground; repeated captures from 2011 to 2021 give each place a temporal stack while the grid is meant to keep neighboring places non-overlapping. Steerable CNNs generate filters at arbitrary orientations as linear combinations of fixed basis filters, making the network equivariant under the cyclic groups $\mathrm{C}_4$ (90 degrees) and $\mathrm{C}_8$ (45 degrees), and pooling over the orientation dimension produces rotation-invariant descriptors. The dataset provides clean place separation plus temporal diversity; the architecture provides rotation robustness that ordinary CNNs would have to learn from data.

What would settle it

Measure the released LASED coordinates and compute nearest-neighbor distances between sampled locations; if any two centers are closer than 400 meters, the no-overlap guarantee fails, and a model trained after removing those pairs should be compared against the original on LASED-Test.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that structured scale and built-in rotation symmetry are the two levers that unlock aerial vPR. LASED is built so that each sampled location is a unique place: neighboring 400-meter circular footprints do not overlap, and each place has revisit images from 2011 to 2021, giving clean supervision with temporal diversity. The steerable ResNet backbones replace ordinary convolutions with filters that rotate with the input, and orientation pooling turns their equivariant feature maps into invariant descriptors. In the authors' experiments, LASED-trained models outperform AerialVL- and ALTO-trained models across all test datasets, and the best steerable model, sresnet50c8, beats the best non-steerable architecture by 12% average Recall@1. The paper also finds that steerable networks still need rotation augmentation to reach their full accuracy, and that bigger backbones only pay off when trained on a dataset as large as LASED.

Load-bearing premise

The dataset's clean place labels depend on the unreported latitude/longitude grid spacing being large enough that the 400-meter-wide circular images of neighboring places never overlap.

Editorial extensions

If this is right

  • LASED-trained models outperform AerialVL- and ALTO-trained models on every evaluation dataset, so training scale and temporal diversity transfer across countries and terrain types.
  • A steerable ResNet-50 with $\mathrm{C}_8$ symmetry adds 12% average Recall@1 over the best non-steerable network when trained on LASED, and 9% or 35% when trained on AerialVL or ALTO.
  • Larger backbone capacity only produces clear gains when the training set is large and diverse; on smaller datasets it yields minimal or negative returns.
  • Steerable models keep higher recall than conventional ones as descriptor dimensionality shrinks, which lowers the memory needed for storing national-scale place descriptors.
  • Rotation augmentation remains necessary even for equivariant networks, so the two techniques are complementary rather than interchangeable.

Reading between the lines

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

  • If the no-overlap property is verified, LASED's per-place structure could support classification-style or class-balanced training schemes that the current retrieval benchmarks do not exercise.
  • The same recipe—national orthophoto archives, water masking, a constant grid, and circular crops—could produce comparable aerial vPR datasets in other countries, provided the grid interval is disclosed and checked.
  • A direct test of the rotational claim would be to evaluate LASED-trained steerable and standard descriptors on non-orthophoto UAV imagery with known yaw angles; the steerable advantage should grow with yaw spread if the mechanism is rotation equivariance.
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.

Referee Report

3 major / 6 minor

Summary. The paper introduces LASED, a large-scale aerial visual place recognition dataset constructed from Estonian orthophotos, with roughly one million images sampled from over 170,000 locations across a decade of temporal coverage, and a companion evaluation set LASED-Test from Switzerland. It also proposes the use of steerable CNNs for aerial vPR, comparing several standard vPR methods and steerable ResNet variants under multi-similarity training. Experiments across six test sets report consistent Recall@N improvements for models trained on LASED over models trained on AerialVL or ALTO, and for steerable networks over non-steerable baselines, including an average 12% Recall@1 improvement claimed in the abstract.

Significance. If the results hold, the paper addresses a real gap: aerial vPR lacks large, structured, temporally rich training data, and rotation robustness is a genuine issue for UAV imagery. The dataset is potentially valuable, the benchmark spans several external datasets plus a held-out test set from a different country, and the steerable-network direction is sensible. The main claims, however, rest on dataset-design assertions that are not verifiable from the paper and on comparisons that conflate several variables. The central empirical pattern is plausible, but the specific attribution to scale plus diversity, and the specific attribution of the accuracy gain to steerability, need tighter experimental support before the paper is publishable.

major comments (3)
  1. [§III-A, §III-B, §III-C] The no-overlap premise of LASED is load-bearing and unverifiable as written. Section III-A states that locations are sampled with a constant latitude/longitude interval, Section III-B asserts that there is no overlap between neighboring locations, and Section III-C sets the ground diameter of each circular image at approximately 400 m, but the sampling step is never reported. Given about 170,000 locations over Estonia's land area, the mean area per location is roughly 0.26 km2, which corresponds to a mean spacing comparable to the footprint diameter; for an equal angular grid at 59°N achieving this density, the east-west spacing is approximately 0.37 km, which is below the 400 m footprint. Thus the asserted separation cannot be checked from the paper and may fail for at least some neighboring cells. This matters because LASED's stated advantage over AerialVL and ALTO is 'clear place separation,' and overlapping views of the same ground would be mined as negatives during multi-similarity training. Please report the exact latitude/longitude step, the land-area value used, and a computed distribution of pairwise distances between neighboring sampled locations, or revise the dataset-design claim accordingly.
  2. [§V-A.3 and Table II] The comparison that supports the claim 'models trained on LASED outperform those trained on smaller, less diverse datasets' is confounded. In Table II, the LASED, AerialVL, and ALTO training sets differ simultaneously in total image count, geographic extent, temporal span, image source (nadir orthophotos versus drone/aircraft frames), sensor characteristics, and label structure; the paper even notes that the AerialVL and ALTO query/reference structure introduces 'a small amount of noise due to overlap between consecutive queries.' The conclusion that geographic coverage and temporal diversity cause the gains is therefore not uniquely supported. Please add at least one controlled comparison, for example training on a LASED subset matched in size to AerialVL or ALTO, or otherwise isolating scale from diversity and label noise. Without such an experiment, the marginal benefit of the dataset's scale and structure over existing alternatives is identifiable only under a strong assumption.
  3. [§IV-C and §V-B.2, Table II] The 12% average improvement attributed to steerable networks is not supported by a capacity-controlled comparison. In Table II, each steerable model is compared against standard ResNet backbones paired with different aggregators (ConvAP, CosPlace, MixVPR, GeM) or with a custom baseline, so the comparison varies both architecture and aggregation. The only controlled side-by-side is Table III, which uses CosPlace as the sole non-steerable baseline. The paper does not report parameter counts or FLOPs for the steerable backbones, and Section V-C discusses model size without giving these numbers; since steerable filters introduce additional orientation structure, the improvement could partly reflect additional capacity rather than equivariance. Please report parameter counts and FLOPs for every method in Table II and Table III, and include parameter-matched or FLOP-matched standard ResNet baselines for the steerable models.
minor comments (6)
  1. [§II] There is a typo in the sentence 'most are relatively small, with total image number well bellow100, 000'; it should read 'well below 100,000'.
  2. [§III-D] The paper states that fixed geometric transformations are applied to LASED-Test to simulate varying orientations and altitudes, but the specific transformations (rotation angles, scale factors, and any altitude simulation method) are never described. Please specify them so the results can be reproduced.
  3. [§V-C and Fig. 5] The axes and aggregation underlying Fig. 5 are not fully precise: it is unclear whether 'average performance' is the mean Recall@1 across the six test datasets or a different average, and the actual parameter counts for the plotted models are not given in the text or caption. Please clarify both.
  4. [Table II caption] The caption notes that GeM produces 2048-dimensional representations while all other methods produce 512-dimensional ones; this capacity difference should be acknowledged in the comparison and ideally controlled for, since descriptor dimensionality can affect retrieval performance independently of the method.
  5. [§III-B] The dataset name 'LASED-Test' is easily confused with a split of LASED; since it is a separate dataset from Switzerland, consider calling it by a distinct name such as 'LASED-Switzerland' or defining the relationship explicitly at first use.
  6. [Abstract and §V-B.2] The '12% recall improvement' figure is reported without specifying which non-steerable architecture it is relative to, which metric (R@1 versus R@5), and how the average is computed across datasets; please define the exact aggregation in the text where the number first appears.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical benchmark with external test sets; no circular reduction found.

full rationale

The paper is an empirical dataset-and-architecture study rather than a derivation. LASED is constructed from Estonian orthophotos, and the central claims are evaluated by training models on LASED and comparing them against models trained on AerialVL and ALTO under an identical training pipeline, with evaluation on external datasets VPAIR, ALTO, SUES-200, University-1652, and AerialVL, plus the geographically disjoint LASED-Test. The steerable CNN claim is tested against non-steerable baselines under the same training conditions (Table II) and with a controlled ablation on rotation augmentation (Table III). No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no equation defines one claimed result in terms of another. The assertion that LASED places do not overlap is an unverified data-construction assumption, but that is a data-quality/verification concern, not circularity. The only mild issue is that LASED and LASED-Test are created by the same group, but the extensive external benchmarks prevent the headline results from reducing to the dataset itself. Self-citations in the references are contextual and not load-bearing for the paper's main conclusions.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper has no theoretical derivation, so no parameters are fitted to data. The free parameters listed are dataset design choices that directly affect the central claims (place separation, temporal diversity, evaluation difficulty). The axioms are unverified domain assumptions about overlap, georeferencing, and the realism of synthetic rotations.

free parameters (4)
  • Location sampling interval (latitude/longitude step)
    The dataset is generated by 'sampled locations' using a constant interval in latitude and longitude (Sec. III-A). The interval is not reported; it determines the number of places (170k) and whether 400-m-diameter circular footprints overlap. It is a hand-chosen design parameter that directly affects place separation and training label quality.
  • Circular image diameter = 500 pixels / ~400 m ground
    Fixed design choice in Sec. III-C; sets the spatial extent of each place and the altitude range covered, influencing the scale of features and the difficulty of the task.
  • Number of temporal samples per location = up to 10
    Table I lists 10 longitudinal samples per location, but the paper says ~1M images across 170k locations, which would imply an average of ~6 samples per location if all locations are used; the exact per-location distribution is not specified. This choice controls temporal diversity.
  • LASED-Test transformation parameters
    Sec. III-D states 'we apply fixed geometric transformations to LASED-Test' to simulate varying orientations and altitudes, but the exact rotation angles, scale factors, and whether altitude is simulated by downsampling are not given. These choices affect the reported recall numbers.
assumptions (4)
  • domain assumption In aerial views, observation span is limited, so neighboring places separated by small gaps do not overlap.
    Stated in Sec. II ('in aerial views observations span is limited, thus no large gaps are necessary among neighboring places') and used in Sec. III-B to claim that LASED places do not overlap. If the actual spacing is smaller than the footprint, training positives are corrupted.
  • domain assumption The orthophotos from Estonian Land Board and swisstopo are accurately georeferenced and pixel-to-pixel aligned across years.
    Sec. III-B relies on pixel-to-pixel correspondence across time frames to provide clean temporal positives. Inaccuracies in co-registration would degrade training labels.
  • domain assumption Rotating the circular image and center-cropping yields realistic UAV viewpoint variations without introducing interpolation artifacts.
    The data augmentation in Sec. III-B assumes rotated circular crops are representative of UAV roll/yaw changes. The paper acknowledges in Table III that rotation augmentations are needed even for steerable networks, and the claim of 'no interpolation noise' (Sec. III-B) is an overstatement.
  • domain assumption The test transformations applied to LASED-Test are representative of real UAV operating conditions.
    Sec. III-D states fixed geometric transformations simulate 'varying orientations and altitudes' but provides no details or validation against real UAV imagery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Place Recognition for Large-Scale UAV Applications." pith.science (2026). https://pith.science/paper/XMGYRF5Q

@misc{pith2026250715089,
  author       = {Pith},
  title        = {Pith review of: Visual Place Recognition for Large-Scale UAV Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XMGYRF5Q}},
  note         = {Machine review of arXiv:2507.15089}
}
read the original abstract

Visual Place Recognition (vPR) plays a crucial role in Unmanned Aerial Vehicle (UAV) navigation, enabling robust localization across diverse environments. Despite significant advancements, aerial vPR faces unique challenges due to the limited availability of large-scale, high-altitude datasets, which limits model generalization, along with the inherent rotational ambiguity in UAV imagery. To address these challenges, we introduce LASED, a large-scale aerial dataset with approximately one million images, systematically sampled from 170,000 unique locations throughout Estonia over a decade, offering extensive geographic and temporal diversity. Its structured design ensures clear place separation significantly enhancing model training for aerial scenarios. Furthermore, we propose the integration of steerable Convolutional Neural Networks (CNNs) to explicitly handle rotational variance, leveraging their inherent rotational equivariance to produce robust, orientation-invariant feature representations. Our extensive benchmarking demonstrates that models trained on LASED achieve significantly higher recall compared to those trained on smaller, less diverse datasets, highlighting the benefits of extensive geographic coverage and temporal diversity. Moreover, steerable CNNs effectively address rotational ambiguity inherent in aerial imagery, consistently outperforming conventional convolutional architectures, achieving on average 12\% recall improvement over the best-performing non-steerable network. By combining structured, large-scale datasets with rotation-equivariant neural networks, our approach significantly enhances model robustness and generalization for aerial vPR.

Figures

Figures reproduced from arXiv: 2507.15089 by the authors.

Figure 1
Figure 1. Examples of images taken from the same location across different years and seasons. The temporal depth and diversity [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Examples of images from the LASED dataset illustrat [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of feature maps generated by a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of the average Recall@1 and standard de￾viation for all models in Table II across all test datasets, when trained on LASED, AerialVL, and Alto. This highlights the impact of dataset scale and diversity on model generalization and recognition performance. D. …
Figure 5
Figure 5. Figure 5: Average Recall@1 across all test datasets for different backbone sizes. Brighter-colored bars correspond to models trained on LASED, faded-colored bars represent training on AerialVL, while bolder-colored bars indicate models trained on Alto. Models trained on LASED be…
Figure 6
Figure 6. Figure 6: Impact of feature dimensionality on vPR performance [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 52 canonical work pages

  1. [1]

    A Survey on Applications of Unmanned Aerial Vehicles Using Machine Learning,

    K. Teixeira, G. Miguel, H. S. Silva, and F. Madeiro, “A Survey on Applications of Unmanned Aerial Vehicles Using Machine Learning,” IEEE Access , vol. 11, pp. 117 582–117 621, 2023, conference Name: IEEE Access. [Online]. Available: https://ieeexplore.ieee.org/document/ 10287979

    work page 2023

  2. [2]

    Multi-layer map: Augmenting semantic visual memory,

    I. T. Papapetros, V . Balaska, and A. Gasteratos, “Multi-layer map: Augmenting semantic visual memory,” in 2020 International Conference on Unmanned Aircraft Systems (ICUAS) , Sep. 2020, pp. 1206–1212, iSSN: 2575-7296. [Online]. Available: https://ieeexplore. ieee.org/abstract/document/9213923

    work page arXiv 2020

  3. [3]

    Light-weight approach for safe landing in populated areas,

    T. Mitroudas, V . Balaska, A. Psomoulis, and A. Gasteratos, “Light-weight approach for safe landing in populated areas,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), May 2024, pp. 10 027–10 032. [Online]. Available: https: //ieeexplore.ieee.org/abstract/document/10611639

    work page arXiv 2024

  4. [4]

    A compilation of UA V applications for precision agriculture,

    P. Radoglou-Grammatikis, P. Sarigiannidis, T. Lagkas, and I. Moscho- lios, “A compilation of UA V applications for precision agriculture,” Computer Networks, vol. 172, p. 107148, May 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S138912862030116X

    work page 2020

  5. [5]

    Long-Term Visual Localization Revisited,

    C. Toft, W. Maddern, A. Torii, L. Hammarstrand, E. Stenborg, D. Safari, M. Okutomi, M. Pollefeys, J. Sivic, T. Pajdla, F. Kahl, and T. Sattler, “Long-Term Visual Localization Revisited,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 4, pp. 2074–2088, Apr. 2022, conference Name: IEEE Transactions on Pattern Analysis and Mach...

    work page arXiv 2022

  6. [6]

    From Coarse to Fine: Robust Hierarchical Localization at Large Scale,

    P.-E. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From Coarse to Fine: Robust Hierarchical Localization at Large Scale,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: IEEE, Jun. 2019, pp. 12 708–12 717. [Online]. Available: https://ieeexplore.ieee.org/document/8953492/

    work page arXiv 2019

  7. [7]

    InLoc: Indoor Visual Localization with Dense Matching and View Synthesis,

    H. Taira, M. Okutomi, T. Sattler, M. Cimpoi, M. Pollefeys, J. Sivic, T. Pajdla, and A. Torii, “InLoc: Indoor Visual Localization with Dense Matching and View Synthesis,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 43, no. 4, pp. 1293–1307, Apr. 2021, conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence....

    work page arXiv 2021

  8. [8]

    Are Large-Scale 3D Models Really Necessary for Accurate Visual Localization?

    A. Torii, H. Taira, J. Sivic, M. Pollefeys, M. Okutomi, T. Pajdla, and T. Sattler, “Are Large-Scale 3D Models Really Necessary for Accurate Visual Localization?” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 43, no. 3, pp. 814–829, Mar. 2021. [Online]. Available: https://ieeexplore.ieee.org/document/8839843/

    work page arXiv 2021

Show all 58 references
  1. [9]

    The Revisiting Problem in Simultaneous Localization and Mapping: A Survey on Visual Loop Closure Detection,

    K. A. Tsintotas, L. Bampis, and A. Gasteratos, “The Revisiting Problem in Simultaneous Localization and Mapping: A Survey on Visual Loop Closure Detection,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 11, pp. 19 929–19 953, Nov. 2022, conference Name...

    2022

  2. [10]

    Visual Place Recognition: A Survey,

    S. Lowry, N. Sunderhauf, P. Newman, J. J. Leonard, D. Cox, P. Corke, and M. J. Milford, “Visual Place Recognition: A Survey,” IEEE Trans- actions on Robotics , vol. 32, no. 1, pp. 1–19, 2016, publisher: Institute of Electrical and Electronics Engineers Inc

    2016

  3. [11]

    Rethinking Visual Geo- localization for Large-Scale Applications,

    G. Berton, C. Masone, and B. Caputo, “Rethinking Visual Geo- localization for Large-Scale Applications,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . New Orleans, LA, USA: IEEE, Jun. 2022, pp. 4868–4878. [Online]. Available: https://ieeexplor...

    2022

  4. [12]

    GSV-Cities: Toward Appropriate Supervised Visual Place Recognition,

    A. Ali-bey, B. Chaib-draa, and P. Gigu `ere, “GSV-Cities: Toward Appropriate Supervised Visual Place Recognition,” Neurocomputing, vol. 513, pp. 194–203, Nov. 2022, arXiv:2210.10239 [cs]. [Online]. Available: http://arxiv.org/abs/2210.10239

    2022 arXiv

  5. [13]

    Mapillary Street-Level Sequences: A Dataset for Lifelong Place Recognition,

    F. Warburg, S. Hauberg, M. Lopez-Antequera, P. Gargallo, Y . Kuang, and J. Civera, “Mapillary Street-Level Sequences: A Dataset for Lifelong Place Recognition,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Seattle, W A, USA: IEEE, Jun. 2020, ...

    2020

  6. [14]

    NetVLAD: CNN Architecture for Weakly Su- pervised Place Recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “NetVLAD: CNN Architecture for Weakly Su- pervised Place Recognition,” 2016, pp. 5297–5307. [On- line]. Available: https://openaccess.thecvf.com/content cvpr 2016/html/ Arandjelovic NetVLAD CNN Architecture CVPR 20...

    2016

  7. [15]

    AnyLoc: Towards Universal Visual Place Recognition,

    N. Keetha, A. Mishra, J. Karhade, K. M. Jatavallabhula, S. Scherer, M. Krishna, and S. Garg, “AnyLoc: Towards Universal Visual Place Recognition,” IEEE Robotics and Automation Letters , vol. 9, no. 2, pp. 1286–1293, Feb. 2024, conference Name: IEEE Robotics and Automation Lett...

    2024

  8. [16]

    Visual Place Recognition in Changing Environments with Sequence Representations on the Distance-Space Domain,

    I. T. Papapetros, I. Kansizoglou, L. Bampis, and A. Gasteratos, “Visual Place Recognition in Changing Environments with Sequence Representations on the Distance-Space Domain,” Machines, vol. 11, no. 5, p. 558, May 2023, number: 5 Publisher: Multidisciplinary Digital Publishing...

    2023

  9. [17]

    Semantic-based visual vocabulary for loop closure detection,

    I. T. Papapetros, K. M. Oikonomou, I. Kansizoglou, K. A. Tsintotas, and A. Gasteratos, “Semantic-based visual vocabulary for loop closure detection,” in 2023 IEEE International Conference on Imaging Systems and Techniques (IST) , Oct. 2023, pp. 1–5, iSSN: 2832-4234. [Online]. ...

    2023

  10. [18]

    GeoCluster: Enhancing Visual Place Recognition in Spatial Domain on Aerial Vehicle Platforms,

    C. Chen, M. He, J. Wang, and Z. Meng, “GeoCluster: Enhancing Visual Place Recognition in Spatial Domain on Aerial Vehicle Platforms,” IEEE Robotics and Automation Letters , vol. 9, no. 3, pp. 3013–3020, Mar. 2024, conference Name: IEEE Robotics and JOURNAL OF LATEX CLASS FILES...

    2024

  11. [19]

    The design and use of steerable filters,

    W. Freeman and E. Adelson, “The design and use of steerable filters,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 13, no. 9, pp. 891–906, Sep. 1991, conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence. [Online]. Available: ...

    1991

  12. [20]

    Learning Steerable Filters for Rotation Equivariant CNNs,

    M. Weiler, F. A. Hamprecht, and M. Storath, “Learning Steerable Filters for Rotation Equivariant CNNs,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . Salt Lake City, UT: IEEE, Jun. 2018, pp. 849–858. [Online]. Available: https: //ieeexplore.ieee.org/...

    2018

  13. [21]

    Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age,

    C. Cadena, L. Carlone, H. Carrillo, Y . Latif, D. Scaramuzza, J. Neira, I. Reid, and J. J. Leonard, “Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age,” IEEE Transactions on Robotics , vol. 32, no. 6, pp. 1309–1332, Dec. 2016,...

    2016

  14. [22]

    BRIEF: Binary Robust Independent Elementary Features,

    M. Calonder, V . Lepetit, C. Strecha, and P. Fua, “BRIEF: Binary Robust Independent Elementary Features,” in Computer Vision – ECCV 2010 , K. Daniilidis, P. Maragos, and N. Paragios, Eds. Berlin, Heidelberg: Springer, 2010, pp. 778–792

    2010

  15. [23]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05) , vol. 1, Jun. 2005, pp. 886–893 vol. 1, iSSN: 1063-6919. [Online]. Available: https://ieeexplore.ieee....

    2005

  16. [24]

    CoHOG: A Light-Weight, Compute-Efficient, and Training-Free Visual Place Recognition Technique for Changing Environments,

    M. Zaffar, S. Ehsan, M. Milford, and K. McDonald-Maier, “CoHOG: A Light-Weight, Compute-Efficient, and Training-Free Visual Place Recognition Technique for Changing Environments,” IEEE Robotics and Automation Letters , vol. 5, no. 2, pp. 1835–1842, Apr. 2020, conference Name: ...

    2020

  17. [25]

    Visual Loop-Closure Detection via Prominent Feature Tracking,

    I. T. Papapetros, V . Balaska, and A. Gasteratos, “Visual Loop-Closure Detection via Prominent Feature Tracking,” Journal of Intelligent & Robotic Systems , vol. 104, no. 3, p. 54, Mar. 2022. [Online]. Available: https://doi.org/10.1007/s10846-022-01581-9

    2022 doi

  18. [26]

    A Survey on Deep Visual Place Recognition,

    C. Masone and B. Caputo, “A Survey on Deep Visual Place Recognition,” IEEE Access, vol. 9, pp. 19 516–19 547, 2021, conference Name: IEEE Access. [Online]. Available: https://ieeexplore.ieee.org/ document/9336674

    2021

  19. [27]

    Recent advances in convolutional neural networks,

    J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang, G. Wang, J. Cai, and T. Chen, “Recent advances in convolutional neural networks,” Pattern Recognition , vol. 77, pp. 354–377, May 2018. [Online]. Available: https: //www.sciencedirect.com/science/article/...

    2018

  20. [28]

    Evaluating the Performance of Mobile-Convolutional Neural Networks for Spatial and Temporal Human Action Recognition Analysis,

    S. N. Moutsis, K. A. Tsintotas, I. Kansizoglou, and A. Gasteratos, “Evaluating the Performance of Mobile-Convolutional Neural Networks for Spatial and Temporal Human Action Recognition Analysis,” Robotics, vol. 12, no. 6, p. 167, Dec. 2023, number: 6 Publisher: Multidisciplina...

    2023

  21. [29]

    ImageNet Classification with Deep Convolutional Neural Networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances in Neural Information Processing Systems , vol. 25. Curran Associates, Inc.,

  22. [30]

    Place recognition with ConvNet landmarks: Viewpoint-robust, condition-robust, training- free,

    N. Suenderhauf, S. Shirazi, A. Jacobson, F. Dayoub, E. Pepperell, B. Upcroft, and M. Milford, “Place recognition with ConvNet landmarks: Viewpoint-robust, condition-robust, training- free,” in Robotics: Science and Systems XI , D. Hsu, Ed. http://www.roboticsproceedings.org/: ...

    2015

  23. [31]

    EigenPlaces: Training Viewpoint Robust Models for Visual Place Recognition,

    G. Berton, G. Trivigno, B. Caputo, and C. Masone, “EigenPlaces: Training Viewpoint Robust Models for Visual Place Recognition,” 2023, pp. 11 080–11 090. [Online]. Available: https://openaccess.thecvf.com/content/ICCV2023/html/ Berton EigenPlaces Training Viewpoint Robust Model...

    2023

  24. [32]

    Generalized Contrastive Optimization of Siamese Networks for Place Recognition,

    M. Leyva-Vallina, N. Strisciuglio, and N. Petkov, “Generalized Contrastive Optimization of Siamese Networks for Place Recognition,” Apr. 2023, arXiv:2103.06638 [cs]. [Online]. Available: http://arxiv.org/ abs/2103.06638

    2023 arXiv

  25. [33]

    VPAIR – Aerial Visual Place Recognition and Localization in Large-scale Outdoor Environments,

    M. Schleiss, F. Rouatbi, and D. Cremers, “VPAIR – Aerial Visual Place Recognition and Localization in Large-scale Outdoor Environments,” May 2022, arXiv:2205.11567 [cs]. [Online]. Available: http://arxiv.org/abs/2205.11567

    2022 arXiv

  26. [34]

    ALTO: A Large-Scale Dataset for UA V Visual Place Recognition and Localization,

    I. Cisneros, P. Yin, J. Zhang, H. Choset, and S. Scherer, “ALTO: A Large-Scale Dataset for UA V Visual Place Recognition and Localization,” Jul. 2022, arXiv:2207.12317 [cs]. [Online]. Available: http://arxiv.org/abs/2207.12317

    2022 arXiv

  27. [35]

    SUES-200: A Multi-Height Multi-Scene Cross-View Image Benchmark Across Drone and Satellite,

    R. Zhu, L. Yin, M. Yang, F. Wu, Y . Yang, and W. Hu, “SUES-200: A Multi-Height Multi-Scene Cross-View Image Benchmark Across Drone and Satellite,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 33, no. 9, pp. 4825–4839, Sep. 2023, conference Name: IEEE T...

    2023

  28. [36]

    AerialVL: A Dataset, Baseline and Algorithm Framework for Aerial- Based Visual Localization With Reference Map,

    M. He, C. Chen, J. Liu, C. Li, X. Lyu, G. Huang, and Z. Meng, “AerialVL: A Dataset, Baseline and Algorithm Framework for Aerial- Based Visual Localization With Reference Map,” IEEE Robotics and Automation Letters , vol. 9, no. 10, pp. 8210–8217, Oct. 2024, conference Name: IEE...

    2024

  29. [37]

    University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization,

    Z. Zheng, Y . Wei, and Y . Yang, “University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization,” Aug. 2020, arXiv:2002.12186 [cs]. [Online]. Available: http://arxiv.org/abs/2002. 12186

    2020 arXiv

  30. [38]

    UA V-VisLoc: A Large-scale Dataset for UA V Visual Localization,

    W. Xu, Y . Yao, J. Cao, Z. Wei, C. Liu, J. Wang, and M. Peng, “UA V-VisLoc: A Large-scale Dataset for UA V Visual Localization,” May 2024, arXiv:2405.11936 [cs]. [Online]. Available: http://arxiv.org/ abs/2405.11936

    2024 arXiv

  31. [39]

    Vision- Based UA V Self-Positioning in Low-Altitude Urban Environments,

    M. Dai, E. Zheng, Z. Feng, L. Qi, J. Zhuang, and W. Yang, “Vision- Based UA V Self-Positioning in Low-Altitude Urban Environments,” IEEE Transactions on Image Processing , vol. 33, pp. 493–508, 2024, conference Name: IEEE Transactions on Image Processing. [Online]. Available: ...

    2024

  32. [40]

    Are State-of-the-art Visual Place Recognition Techniques any Good for Aerial Robotics?

    M. Zaffar, A. Khaliq, S. Ehsan, M. Milford, K. Alexis, and K. McDonald-Maier, “Are State-of-the-art Visual Place Recognition Techniques any Good for Aerial Robotics?” May 2019, arXiv:1904.07967 [cs]. [Online]. Available: http: //arxiv.org/abs/1904.07967

    2019 arXiv

  33. [41]

    Visual place recognition for aerial imagery: A survey,

    I. Moskalenko, A. Kornilova, and G. Ferrer, “Visual place recognition for aerial imagery: A survey,” Robotics and Autonomous Systems , vol. 183, p. 104837, Jan. 2025. [Online]. Available: https://www. sciencedirect.com/science/article/pii/S0921889024002215

    2025

  34. [42]

    Delta Descriptors: Change-Based Place Representation for Robust Visual Localization,

    S. Garg, B. Harwood, G. Anand, and M. Milford, “Delta Descriptors: Change-Based Place Representation for Robust Visual Localization,” IEEE Robotics and Automation Letters , vol. 5, no. 4, pp. 5120–5127, 2020, conference Name: IEEE Robotics and Automation Letters

    2020

  35. [43]

    Deep Feature Space: A Geometrical Perspective,

    I. Kansizoglou, L. Bampis, and A. Gasteratos, “Deep Feature Space: A Geometrical Perspective,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 10, pp. 6823–6838, Oct. 2022, conference Name: IEEE Transactions on Pattern Analysis and Machine Intelli...

    2022

  36. [44]

    Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning,

    X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott, “Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Long Beach, CA, USA: IEEE, Jun. 2019, pp. 5017–5025. [Online]. Ava...

    2019

  37. [45]

    Learning local feature descriptors with triplets and shallow convolutional neural networks,

    V . Balntas, E. Riba, D. Ponsa, and K. Mikolajczyk, “Learning local feature descriptors with triplets and shallow convolutional neural networks,” in Procedings of the British Machine Vision Conference

  38. [46]

    CosFace: Large Margin Cosine Loss for Deep Face Recognition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu, “CosFace: Large Margin Cosine Loss for Deep Face Recognition,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, Jun. 2018, pp. 5265–5274. [Online]. Availabl...

    2018

  39. [47]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016, pp. 770–778, iSSN: 1063-6919. [Online]. Available: https://ieeexplore.ieee.org/document/7780459

    2016

  40. [48]

    Very Deep Convolutional Networks for Large-Scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Apr. 2015, arXiv:1409.1556 [cs]. [Online]. Available: http://arxiv.org/abs/1409.1556

    2015 arXiv

  41. [49]

    Fine-Tuning CNN Image Retrieval with No Human Annotation,

    F. Radenovi ´c, G. Tolias, and O. Chum, “Fine-Tuning CNN Image Retrieval with No Human Annotation,” IEEE Transactions on JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, X XXXX 12 Pattern Analysis and Machine Intelligence , vol. 41, no. 7, pp. 1655–1668, Jul. 2019, conference Nam...

    2019

  42. [50]

    A Transformer-Based Feature Segmentation and Region Alignment Method For UA V-View Geo- Localization,

    M. Dai, J. Hu, J. Zhuang, and E. Zheng, “A Transformer-Based Feature Segmentation and Region Alignment Method For UA V-View Geo- Localization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 7, pp. 4376–4389, Jul. 2022, arXiv:2201.09206 [cs]. [Onl...

    2022 arXiv

  43. [51]

    Each Part Matters: Local Patterns Facilitate Cross-view Geo-localization,

    T. Wang, Z. Zheng, C. Yan, J. Zhang, Y . Sun, B. Zheng, and Y . Yang, “Each Part Matters: Local Patterns Facilitate Cross-view Geo-localization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 2, pp. 867–879, Feb. 2022, arXiv:2008.11646 [cs]. [Onl...

    2022 arXiv

  44. [52]

    Google Earth Engine: Planetary-scale geospatial analysis for everyone,

    N. Gorelick, M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, and R. Moore, “Google Earth Engine: Planetary-scale geospatial analysis for everyone,” Remote Sensing of Environment , vol. 202, pp. 18–27, Dec. 2017. [Online]. Available: https://www.sciencedirect.com/science/ artic...

    2017

  45. [53]

    High- resolution mapping of global surface water and its long-term changes,

    J.-F. Pekel, A. Cottam, N. Gorelick, and A. S. Belward, “High- resolution mapping of global surface water and its long-term changes,” Nature, vol. 540, no. 7633, pp. 418–422, Dec. 2016, publisher: Nature Publishing Group. [Online]. Available: https: //www.nature.com/articles/n...

    2016

  46. [54]

    A Program to Build E(N)- Equivariant Steerable CNNs,

    G. Cesa, L. Lang, and M. Weiler, “A Program to Build E(N)- Equivariant Steerable CNNs,” Oct. 2021. [Online]. Available: https: //openreview.net/forum?id=WE4qe9xlnQw

    2021

  47. [55]

    MixVPR: Feature Mixing for Visual Place Recognition,

    A. Ali-Bey, B. Chaib-Draa, and P. Giguere, “MixVPR: Feature Mixing for Visual Place Recognition,” in 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) . Waikoloa, HI, USA: IEEE, Jan. 2023, pp. 2997–3006. [Online]. Available: https://ieeexplore.ieee.org/...

    2023

  48. [56]

    RandAugment: Practical automated data augmentation with a reduced search space,

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. V . Le, “RandAugment: Practical automated data augmentation with a reduced search space,” Nov. 2019, arXiv:1909.13719 [cs]. [Online]. Available: http://arxiv.org/abs/1909.13719

    2019 arXiv

  49. [2012]

    Available: https://papers.nips.cc/paper files/paper/2012/ hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html

    [Online]. Available: https://papers.nips.cc/paper files/paper/2012/ hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html

    2012

  50. [2016]

    119.1–119.11

    York, UK: British Machine Vision Association, 2016, pp. 119.1–119.11. [Online]. Available: http://www.bmva.org/bmvc/2016/ papers/paper119/index.html

    2016

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.