Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Hierarchical Image Matching for UAV Absolute Visual Localization via Semantic and Structural Constraints

T0 review · 3 major / 5 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 This paper claims that a hierarchical semantic-plus-structure matching pipeline achieves 0.81 average success rate on AerialVL and 0.93 on a new CS-UAV dataset, locating UAVs from satellite maps without GNSS or relative localization.

desk verdict Solid engineering contribution with a real evaluation-protocol flaw: the reported localization errors depend on a strictly nadir camera and planar-scene assumption that the paper never tests. read the letter →

arxiv 2506.09748 v1 pith:7R3NLPHG submitted 2025-06-11 cs.CV cs.RO

Xiangkai Zhang , Xiang Zhou , Mao Chen , Yuchen Lu , Xu Yang , Zhiyong Liu This is my paper
classification cs.CVcs.RO
keywords UAVvisuallocalizationabsolutecross-sourceimagematchingsemanticfeaturesstructureconstraintscoarse-to-fineGNSS-deniednavigationDINOv2
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 claims that a two-stage image matcher, which first aligns drone images to satellite images using semantic features and structural constraints and then refines the alignment with a lightweight keypoint matcher, locates a UAV in global coordinates without GNSS or relative localization. On the AerialVL benchmark the full pipeline reports a 0.81 success rate and 18.73 m mean localization error, and on a new CS-UAV dataset 0.93 and 17.71 m, compared with 0.36–0.56 success and 24–30 m errors for SIFT, SuperPoint, and Deep-LK baselines run in the same retrieval-plus-matching pipeline. The authors argue that high-level semantic features are more stable than low-level features across the cross-source and temporal gap between UAV video and satellite imagery, and that structural consistency resolves the ambiguity that pure semantic matching would suffer. If correct, the result is a GNSS-free localization pipeline that needs only a downward camera and publicly available satellite maps.

What carries the argument

The load-bearing machinery is the semantic-aware and structure-constrained matching module (SASCM) combined with a lightweight fine-grained matcher. SASCM computes pairwise cosine similarities between DINOv2 dense feature vectors to form a 4D correlation tensor, applies SoftMNN filtering, and runs three 4D convolutional layers to enforce neighborhood consensus, then softmax-normalizes to get probabilistic region-level matches. The fine module, following the XFeat architecture, detects keypoints via reliability maps and 8x8 cell classification, matches them mutually, and estimates a homography $H$ with RANSAC; the UAV image center $C_U$ is mapped through $H$ to obtain the position estimate. This coarse-to-fine design is what the paper claims carries the accuracy improvement.

What would settle it

Take a UAV flight with recorded pitch and roll (or over terrain with known elevation changes), run the pipeline, and compare estimated versus ground-truth positions. If localization error grows systematically with camera tilt or terrain relief, the center-correspondence and single-homography assumption is violated and the reported errors understate true performance.

Watch

Extended reading notes

Core claim

The central claim is that UAV-to-satellite image matching for absolute localization is best decomposed into a coarse semantic stage and a fine texture stage, bridged by region-level correspondences. The coarse module (SASCM) extracts dense DINOv2 features from both images, builds a 4D similarity tensor, filters ambiguous matches with soft mutual-nearest-neighbor and 4D convolutional neighborhood consensus, and returns a probability distribution over matches; the fine module then detects keypoints and descriptors only inside the matched central region, using mutual nearest neighbor and RANSAC to estimate a homography that maps the UAV image center to the satellite image. Because the camera is assumed to face straight down, that mapped point is the UAV's geographic position. The paper reports that this pipeline, which does not use any relative localization technique, outperforms classical and learned low-level matchers integrated in the same pipeline, and also beats the reported AerialVL combined baselines that do use relative localization.

Load-bearing premise

The whole localization step rests on the assumptions that the UAV camera is strictly downward-facing and the scene is locally planar, so the center of the UAV image maps through one homography to the UAV's true ground position.

Editorial extensions

If this is right

  • On AerialVL, the pipeline reports an average success rate of 0.81 and mean localization error of 18.73 m, with per-trajectory success rates from 0.53 to 0.96.
  • On CS-UAV, it reports 0.93 average success and 17.71 m error, beating SIFT, SuperPoint, and Deep-LK integrated into the same retrieval pipeline.
  • Ablation shows retrieval-only drops to 0.41/0.54 success, and removing SASCM drops to 0.68/0.72, so both stages contribute to the reported performance.
  • The pipeline needs no relative localization (no odometry or INS), only a downward-facing camera and a satellite map of the area.

Reading between the lines

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

  • A natural extension is to relax the strict downward-facing assumption by estimating the homography for a tilted camera or fusing an IMU prior, which could extend the method to oblique views.
  • The semantic stage may transfer to other cross-source matching tasks, such as aligning historical maps or ground-level photos with satellite imagery, where appearance differs but scene structure persists.
  • The reported numbers are benchmark-specific; the relative gain over low-level matchers suggests the bottleneck in GNSS-denied localization is semantic correspondence rather than keypoint precision.
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 / 5 minor

Summary. The paper proposes a hierarchical pipeline for UAV absolute visual localization from UAV imagery to satellite reference maps. The pipeline consists of an image retrieval stage using DINOv2 features with optimal transport aggregation, a semantic-aware and structure-constrained coarse matching module (SASCM) that builds a 4D correlation tensor and applies neighborhood consensus filtering, and a lightweight fine-grained matching module based on convolutional keypoint detection and description. The final position is obtained by mapping the UAV image center through a RANSAC-estimated homography. Experiments are reported on the AerialVL benchmark and a newly introduced CS-UAV dataset, comparing the full pipeline against SIFT, SuperPoint, and Deep-LK integrated into the same retrieval and localization framework, alongside two retrieval-plus-matching baselines from the AerialVL paper. The authors report average success rates of 0.81 (AerialVL) and 0.93 (CS-UAV) with mean localization errors of 18.73 m and 17.71 m, respectively, outperforming the integrated baselines.

Significance. If the reported results are accurate, the paper makes a useful empirical contribution to UAV visual localization in GNSS-denied scenarios. The module-level comparisons against SIFT, SuperPoint, and Deep-LK within the same pipeline are a strength, as is the introduction of a new evaluation dataset, CS-UAV, which covers diverse terrain including mountainous forests. The use of semantic features from a vision foundation model for coarse matching is a sensible and timely idea, and the ablations indicate that both the semantic matching stage and the fine matching stage contribute to the reported performance. However, the evaluation protocol and the geometric assumptions underlying the final localization step are not sufficiently validated; these issues are load-bearing for the central claim of superior accuracy, because they may bias the reported metrics in the method's favor and because the measured quantity may not be the UAV's true ground position under realistic camera tilt or terrain relief.

major comments (3)
  1. [Section IV-C, Tables I and II] The evaluation protocol states that 'any error exceeding 50 meters is regarded as drift' and that 'success rate and mean localization error [are computed] within the non-drift segments.' This censoring removes the worst errors from both the numerator and denominator of the success rate and from the mean error computation. Since different methods may drift at very different rates, the reported numbers do not compare overall localization performance; a method that drifts on 30% of frames and a method that drifts on 5% of frames are compared only on their non-drift subsets. The paper should report drift rates per trajectory and uncensored statistics (e.g., the fraction of frames below 25 m and 50 m, and the median or trimmed error over all frames) so that the reader can judge whether the reported superiority of the proposed method is driven by a lower drift rate or by better accuracy on successful segments.
  2. [Section III-B and III-C] The final position estimate is obtained by mapping the UAV image center C_U through the homography H, justified by the assumption that the UAV camera is strictly downward-facing and that the scene is locally planar. The paper does not report camera attitude or terrain slope for either dataset, and the CS-UAV dataset explicitly includes mountainous forests. Under a tilt of theta, the offset between the optical-axis ground intersection and the true nadir projection is approximately h * tan(theta), which is 8.7 m at h = 100 m and theta = 5 degrees, and 17.6 m at theta = 10 degrees. These values are comparable to the 25 m success threshold and to the reported mean errors of 18.73 m and 17.71 m. The paper should either quantify the sensitivity of the reported errors to plausible tilt and terrain relief, report attitude and terrain statistics for the evaluation sequences, or explicitly scope the claim to strictly nadir, planar scenes. As written, the central experimental claim may be measuring error relative to the optical-axis intersection rather than the UAV's true ground position.
  3. [Table I and Section IV-C] The text says the comparison against the AerialVL baselines [19] is performed 'across three representative flight trajectories,' but Table I reports results for eleven trajectories, and the AerialVL Comb.1 and Comb.2 rows contain missing entries for many of them. It is unclear which trajectories are actually used in the comparison and whether the averages are computed over different subsets for different methods, which would make the comparison unfair. The paper should specify exactly which sequences are used for each baseline, report the missing entries (or state clearly that the original paper did not provide them), and avoid computing or implying an aggregate comparison over an inconsistent set of trajectories.
minor comments (5)
  1. [Equation (10)] The loss definition appears to contain a typo: it reads 'Mean(fMU) + Mean(fMU)', but the second term should presumably be Mean(fMS), since the softmax is applied along both the UAV and satellite dimensions in Eqs. (8) and (9).
  2. [Equation (5)] The notation '[ReLU(4D-Conv(S))] × 3' is ambiguous; it should be clarified whether this denotes three sequential 4D convolution layers with ReLU activations, and the channel dimensions of the 4D kernels should be specified.
  3. [Section IV-A] The CS-UAV dataset description states that the dataset is 'publicly accessible at [link],' which is a placeholder. The link should be provided in the final version so that the new dataset can actually be used by the community.
  4. [Table I] There is a formatting issue in the Deep-LK row under ST.4/ST.5: '0.530.48' appears without a space or separator, and the table layout should be cleaned up for readability.
  5. [Section IV-C, Ablation Study] The retrieval-only ablation on CS-UAV reports an MLE of 24.20 m and the SIFT baseline reports 24.21 m, which are nearly identical; the text should explain why the retrieval-only result is not substantially worse than SIFT on this dataset, since the two methods are conceptually different (retrieval gives a coarse patch center, while SIFT performs geometric matching).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported results are measured on external datasets against GPS ground truth, and no load-bearing step reduces to a fitted parameter, a self-citation, or a definitional tautology.

full rationale

The paper is an empirical systems paper, not a derivation from first principles. The central claims (success rate 0.81 / MLE 18.73 m on AerialVL; 0.93 / 17.71 m on CS-UAV) are measured by comparing the pipeline's output position against onboard high-precision GPS ground truth on external evaluation datasets (Section IV-A, Tables I-II). The coarse matching network is trained on UAV-Visloc with positive pairs formed from GPS-labeled satellite patches and negative pairs from random non-corresponding regions, and the fine matching network is trained on synthetic geometric/photometric transformations with known pixel correspondences (Section IV-B); neither training signal is derived from the method's own outputs. The nadir/planar assumption in Section III-B, stated as 'Based on the previous assumption that the UAV camera maintains a strictly downward-facing orientation, the central region of the UAV image can be regarded as corresponding to the actual geographic location of the UAV,' is a geometric modeling assumption about the sensor, not a tautological reduction: the resulting position estimate is compared with independent GPS measurements, so the assumption is empirically falsifiable. All cited components (DINOv2, optimal-transport retrieval, XFeat, NCNet-style weak supervision, UAV-Visloc, AerialVL) are external prior work; none of the 26 references is authored by the present authors, and no load-bearing 'uniqueness' or 'prediction' claim rests on a self-citation. The metric protocol (success = error < 25 m, MLE computed over non-drift segments) is inherited from the external AerialVL benchmark and applied identically to baselines; it is a measurement convention, not a circular construction. No equation in the paper reduces a claimed prediction to a fitted parameter or to the definition of the output by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method relies on a family of domain assumptions (nadir camera, planar scene, synthetic-to-real transfer, DINOv2 transferability, retrieval reliability) rather than a small set of fitted constants. Two free parameters (keypoint threshold and crop sizes) are unreported. No new physical entities are introduced; the CS-UAV dataset is a data contribution, not a postulated entity. These assumptions are load-bearing because the conversion from image matches to geographic coordinates depends on them.

free parameters (2)
  • Keypoint extraction threshold sigma = not reported
    Threshold for selecting keypoints in the fine matching module (e^{i_j} * e^{r_i_j} > sigma); controls keypoint density and matching quality, but its value is not disclosed.
  • Crop sizes for central region A and satellite region B = not reported
    Sizes of the image patches passed from the coarse match to the fine matcher are not given; they trade off accuracy and runtime.
assumptions (5)
  • domain assumption The UAV camera is strictly downward-facing, so the center of the UAV image corresponds to the true geographic location of the UAV.
    Stated in Section III-B and used to convert the coarse match into a geographic position; violated by any camera tilt.
  • domain assumption The scene is locally planar so a homography maps the UAV image to the satellite image.
    Used in Section III-C to map the UAV image center to the satellite map with RANSAC homography; weak for terrain with strong relief.
  • domain assumption DINOv2 semantic features provide correspondences that are robust to UAV-satellite cross-source and temporal differences.
    The success of the coarse SASCM module depends on the transferability of DINOv2 features across domains; not explicitly validated separately.
  • domain assumption Synthetic geometric and photometric transformations of UAV images are sufficient to train the fine matching network to generalize to real satellite imagery.
    The fine matcher is trained only on transformed UAV images (Section IV-B); no validation of this synthetic-to-real transfer is given.
  • domain assumption The off-the-shelf image retrieval model retrieves a satellite image that overlaps the UAV's true location.
    The retrieval module [18] is used as a black box; retrieval failure would invalidate the downstream matching, but no retrieval accuracy or failure analysis is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Image Matching for UAV Absolute Visual Localization via Semantic and Structural Constraints." pith.science (2026). https://pith.science/paper/7R3NLPHG

@misc{pith2026250609748,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Image Matching for UAV Absolute Visual Localization via Semantic and Structural Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7R3NLPHG}},
  note         = {Machine review of arXiv:2506.09748}
}
read the original abstract

Absolute localization, aiming to determine an agent's location with respect to a global reference, is crucial for unmanned aerial vehicles (UAVs) in various applications, but it becomes challenging when global navigation satellite system (GNSS) signals are unavailable. Vision-based absolute localization methods, which locate the current view of the UAV in a reference satellite map to estimate its position, have become popular in GNSS-denied scenarios. However, existing methods mostly rely on traditional and low-level image matching, suffering from difficulties due to significant differences introduced by cross-source discrepancies and temporal variations. To overcome these limitations, in this paper, we introduce a hierarchical cross-source image matching method designed for UAV absolute localization, which integrates a semantic-aware and structure-constrained coarse matching module with a lightweight fine-grained matching module. Specifically, in the coarse matching module, semantic features derived from a vision foundation model first establish region-level correspondences under semantic and structural constraints. Then, the fine-grained matching module is applied to extract fine features and establish pixel-level correspondences. Building upon this, a UAV absolute visual localization pipeline is constructed without any reliance on relative localization techniques, mainly by employing an image retrieval module before the proposed hierarchical image matching modules. Experimental evaluations on public benchmark datasets and a newly introduced CS-UAV dataset demonstrate superior accuracy and robustness of the proposed method under various challenging conditions, confirming its effectiveness.

Figures

Figures reproduced from arXiv: 2506.09748 by the authors.

Figure 1
Figure 1. An overview of the UAV absolute visual localization task, in which [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed UAV absolute visual localization pipeline under GNSS-denied conditions. The pipeline consists of three key components: [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Diagram of the image retrieval module, where the green arrows [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Fig.4 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 4
Figure 4. Figure 4: Architecture of the proposed Semantic-Aware and Structure-Constrained Matching Module (SASCM). This module first extracts dense semantic [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Lightweight fine-grained matching module based on a convolu [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The top image illustrates several example UAV flight trajectories [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SASGeo: Stability-Aware Semantic Map Localization for GNSS-Denied UAVs -- A Framework and Synthetic Proof of Concept

    cs.RO 2026-07 conditional novelty 4.5 of 10

    Spatial semantic raster matching retrieves the correct map location at 94.5–95.5% Recall@1 under hard synthetic cross-view perturbations, versus 58.6% for a global semantic descriptor.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [19]

    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, 2024

    work page 2024

  2. [1]

    Analysis of military application of UA V swarm technology,

    X. Zhu, “Analysis of military application of UA V swarm technology,” inInternational Conference on Unmanned Systems, 2020, pp. 1200- 1204

    work page 2020

  3. [2]

    Edge Intelligence in Intelligent Transportation Systems: A Survey,

    T. Gong, L. Zhu, F. R. Yu, and T. Tang, “Edge Intelligence in Intelligent Transportation Systems: A Survey,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 9, pp. 8919-8944, 2023

    work page 2023

  4. [3]

    An autonomous multi-UA V system for search and rescue,

    J. Scherer, S. Yahyanejad, S. Hayat, E. Yanmaz, T. Andre, A. Khan, V . Vukadinovic, C. Bettstetter, H. Hellwagner, and B. Rinner, “An autonomous multi-UA V system for search and rescue,” inProceedings of the First Workshop on Micro Aerial V ehicle Networks, Systems, and Applications for Civilian Use, 2015, pp. 33–38

    work page 2015

  5. [4]

    Search and rescue rotary-wing UA V and its application to the Lushan MS 7.0 earthquake,

    J. Qi, D. Song, H. Shang, N. Wang, C. Hua, C. Wu, X. Qi, and J. Han, “Search and rescue rotary-wing UA V and its application to the Lushan MS 7.0 earthquake,”Journal of Field Robotics, vol. 33, no. 3, pp. 290–321, 2016

    work page 2016

  6. [5]

    Exploring Deep Learning-Based Visual Localization Techniques for UA Vs in GPS-Denied Environments,

    O. Y . Al-Jarrah, A. S. Shatnawi, M. M. Shurman, O. A. Ramadan, and S. Muhaidat, “Exploring Deep Learning-Based Visual Localization Techniques for UA Vs in GPS-Denied Environments,”IEEE Access, vol. 12, pp. 113049-113071, 2024

    work page 2024

  7. [6]

    A review on absolute visual localization for UA V ,

    A. Couturier and M. A. Akhloufi, “A review on absolute visual localization for UA V ,”Robotics and Autonomous Systems, vol. 135, p. 103666, 2021

    work page 2021

  8. [7]

    Leveraging map retrieval and alignment for robust UA V visual geo-localization,

    M. He, J. Liu, P. Gu, and Z. Meng, “Leveraging map retrieval and alignment for robust UA V visual geo-localization,”IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1-13, 2024

    work page 2024

Show all 26 references
  1. [8]

    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,” inProceedings of the 28th ACM International Conference on Multimedia, 2020, pp. 1395–1403

    2020

  2. [9]

    UA V-Satellite View Synthesis for Cross-View Geo-Localization,

    X. Tian, J. Shao, D. Ouyang, and H. T. Shen, “UA V-Satellite View Synthesis for Cross-View Geo-Localization,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 7, pp. 4804– 4815, 2022

    2022

  3. [10]

    A scene graph encoding and matching network for UA V visual localization,

    R. Duan, L. Chen, Z. Li, Z. Chen, and B. Wu, “A scene graph encoding and matching network for UA V visual localization,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 9890–9902, 2024

    2024

  4. [11]

    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, 2023

    2023

  5. [12]

    A deep CNN-based framework for enhanced aerial imagery registration with applications to UA V geolocalization,

    A. Nassar, K. Amer, R. ElHakim, and M. ElHelw, “A deep CNN-based framework for enhanced aerial imagery registration with applications to UA V geolocalization,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 1513– 1523

    2018

  6. [13]

    GPS-denied UA V localization using pre- existing satellite imagery,

    H. Goforth and S. Lucey, “GPS-denied UA V localization using pre- existing satellite imagery,” inProceedings of IEEE International Conference on Robotics and Automation, 2019, pp. 2974–2980

    2019

  7. [14]

    Lucas-Kanade 20 years on: A unifying framework,

    S. Baker and I. Matthews, “Lucas-Kanade 20 years on: A unifying framework,”International Journal of Computer Vision, vol. 56, pp. 221–255, 2004

    2004

  8. [15]

    Assisting UA V localization via deep contextual image matching,

    M. H. Mughal, M. J. Khokhar, and M. Shahzad, “Assisting UA V localization via deep contextual image matching,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 2445–2457, 2021

    2021

  9. [16]

    A coarse-to-fine visual geo-localization method for GNSS-denied UA V with oblique-view imagery,

    Q. Ye, J. Luo, and Y . Lin, “A coarse-to-fine visual geo-localization method for GNSS-denied UA V with oblique-view imagery,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 212, pp. 306– 322, 2024

    2024

  10. [17]

    Leveraging semantic cues from foundation vision models for enhanced local feature correspondence,

    F. Cadar, G. Potje, R. Martins, C. Demonceaux, and E. R. Nasci- mento, “Leveraging semantic cues from foundation vision models for enhanced local feature correspondence,” inProceedings of the Asian Conference on Computer Vision, 2025, pp. 50–66

    2025

  11. [18]

    Optimal transport aggregation for visual place recognition,

    S. Izquierdo and J. Civera, “Optimal transport aggregation for visual place recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17658–17668

    2024

  12. [20]

    SuperPoint: Self- supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperPoint: Self- supervised interest point detection and description,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 224–236

    2018

  13. [21]

    Semantic image matching for cross-domain image retrieval,

    L. Zhang and L. Liu, “Semantic image matching for cross-domain image retrieval,”IEEE Transactions on Image Processing, vol. 30, pp. 3894-3907, 2021

    2021

  14. [22]

    XFeat: Accelerated features for lightweight image matching,

    G. Potje, F. Cadar, A. Araujo, R. Martins, and E. R. Nascimento, “XFeat: Accelerated features for lightweight image matching,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 2682–2691

    2024

  15. [23]

    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,”arXiv preprint arXiv:2405.11936, 2024

    2024 arXiv

  16. [24]

    Distinctive Image Features from Scale-Invariant Key- points,

    D. G. Lowe, “Distinctive Image Features from Scale-Invariant Key- points,”International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004

    2004

  17. [25]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. Yahia, M. Szafraniec, V . El Kholy, M. Henaff, W. Galuba, E. Kharitonov, I. Misra, H. Jegou, and J. Carreira, “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  18. [26]

    NCNet: Neighbourhood consensus networks for estimating image correspondences,

    I. Rocco, M. Cimpoi, R. Arandjelovi ´c, A. Torii, T. Pajdla, and J. Sivic, “NCNet: Neighbourhood consensus networks for estimating image correspondences,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 2, pp. 1020–1034, 2020

    2020

Pith tools

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