Pith. sign in

REVIEW 5 major objections 6 minor 34 references

JRN-Geo: A Joint Perception Network based on RGB and Normal images for Cross-view Geo-localization

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding monocular surface normals to RGB and fusing the two through a difference-aware dual-branch network achieves the best published cross-view geo-localization accuracy on University-1652 and SUES-200.

desk verdict A credible RGB+normal fusion recipe for cross-view geo-localization with real gains on two benchmarks, but the paper never validates that the normal maps carry the geometric signal it claims. read the letter →

arxiv 2509.05696 v1 pith:ZBYBXXVG submitted 2025-09-06 cs.CV

Hongyu Zhou , Yunzhou Zhang , Tingsong Huang , Fawei Ge , Man Qi , Xichen Zhang , Yizhong Zhang This is my paper · ORCID
classification cs.CV
keywords cross-viewgeo-localizationsurfacenormalestimationmulti-modalfeaturefusionviewpoint-invariantdescriptorsUAVlocalizationdual-branchnetwork3Dgeographicaugmentationdrone-satelliteretrieval
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 is trying to establish that geometric structure, in the form of surface-normal maps estimated from a single image, closes part of the drone-to-satellite viewpoint gap that RGB-only methods leave open. It builds a dual-branch network that fuses color features with normal-map features, using a difference-aware module that treats the RGB-normal discrepancy as a signal to amplify, and a cross-branch aggregation step that lets the two modalities constrain each other. It also introduces a 3D reconstruction-based augmentation that crops new drone-satellite instance pairs from existing data, so the network sees more viewpoint variation without new annotations. On the University-1652 and SUES-200 benchmarks the method reports top accuracy in both retrieval directions and at every tested altitude—up to 99.28% Recall@1—with the largest gains at the lowest altitude, where viewpoint differences are most severe. If the mechanism is what the numbers suggest, normal maps give cross-view localization a genuinely geometric matching cue rather than another appearance-based one.

What carries the argument

The load-bearing object is the normal-image modality: per-pixel surface orientation from a monocular estimator (DPT-Hybrid, trained largely on indoor scans), which the paper argues is stable under lighting and viewpoint shifts in a way color is not. Fusion rests on the identity f_r = (f_r+f_n)/2 + (f_r−f_n)/2 (and symmetrically for f_n), splitting features into shared common-mode and differential RGB-minus-normal content; an activation δ(x)=1−4·sigmoid′(x) weights the difference so the network emphasizes what the normal branch adds beyond RGB. Cross-branch spatial attention then aggregates the two streams into joint-constrained descriptors. The third mechanism is data-level: COLMAP reconstru

What would settle it

Replace the normal branch's inputs with spatially shuffled or left-right mirrored normal maps and rerun the University-1652 experiment: if Recall@1 stays nearly unchanged, the geometric content of the normals is not driving the reported gains. Alternatively, render ground-truth normals from the COLMAP reconstruction used for the augmentation and measure the angular error of the estimated maps on drone and satellite views—large errors would mean the structural cue is unreliable across viewpoints.

Watch

Extended reading notes

Core claim

The paper's central claim, on its own terms, is that surface-normal maps—per-pixel estimates of 3D surface orientation—carry structure that survives the drone-to-satellite viewpoint change better than color and texture do, and that a network built to fuse both exploits this. Each RGB/normal feature pair is decomposed into common and difference components, the difference is amplified by a sigmoid-derived activation, and the branches are recombined under joint spatial constraints; at the data level, COLMAP reconstructions are used to crop new matched drone-satellite pairs from existing images, multiplying training viewpoints without new annotations. The paper reports best-published Recall@1 an

Load-bearing premise

The whole approach assumes that a monocular normal estimator trained mostly on indoor scans produces surface-normal maps that are accurate on outdoor drone and satellite imagery and similar enough across the huge viewpoint gap to serve as a shared geometric cue.

Editorial extensions

If this is right

  • Combining estimated surface normals with RGB beats using either modality alone on both benchmarks, in both retrieval directions (drone→satellite and satellite→drone).
  • The differential-mode fusion (DAFM) and cross-branch aggregation (JCIA) each add accuracy over the dual-branch baseline, and their gains combine—evidence that semantic and structural features are complementary rather than redundant.
  • 3D geographic augmentation expands cross-view training data without new annotation, up to a point (k=4) beyond which overlapping crops begin to confuse instance discrimination.
  • Sustained accuracy across drone altitudes from 150m to 300m on SUES-200 implies the structural cue helps most when viewpoint differences are largest.
  • Because the fused descriptor no longer leans primarily on color and texture, matching should hold up better under appearance changes such as season or lighting than RGB-only schemes do.

Reading between the lines

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

  • The paper leaves the cross-view consistency of the normal maps unmeasured: the estimator runs on each image separately and outputs camera-frame normals, so part of the claimed viewpoint-invariant structure may actually be view-dependent. A direct extension is to estimate normals in a ground-aligned (roof-plane) coordinate frame and test whether matching improves further.
  • The common-mode/differential-mode fusion identity is generic: the same DAFM recipe could take any paired modalities—depth, thermal, event frames, infrared—so it is a candidate fusion primitive for retrieval beyond RGB plus normals.
  • The augmentation step currently requires a COLMAP reconstruction and a manually annotated satellite imaging region; automating that annotation would let the technique scale from instance-level benchmarks to city-wide image collections.
  • If geometric content is truly what carries the gain, the same architecture should transfer to cross-view matching across seasonal or day/night appearance changes with little retraining—a regime neither of the two benchmarks exercises.
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

5 major / 6 minor

Summary. The paper proposes JRN-Geo, a dual-branch network for cross-view geo-localization that fuses RGB images with monocular surface-normal images. A Difference-Aware Fusion Module (DAFM) and a Joint-Constrained Interaction Aggregation (JCIA) strategy are introduced to combine semantic and structural features. The authors also propose a 3D geographic augmentation technique based on COLMAP to mine additional cross-view training pairs. Experiments on University-1652 and SUES-200 report state-of-the-art Recall@1 and AP, with ablations showing improvements from each component.

Significance. The idea of using normal maps as a geometric cue in cross-view geo-localization is timely and could be a useful complement to RGB-only features. The paper provides component ablations (Tables III and IV) that support the individual contributions, and the 3D augmentation approach is conceptually interesting. However, the reported SOTA numbers are not fully supported because the geometric mechanism is not validated, hyperparameter k is tuned on the test benchmarks, and Table II contains apparent data errors. If these issues are addressed, the method could be a solid contribution. The manuscript is not currently reproducible due to ambiguous equations and missing code.

major comments (5)
  1. [§III-B/§IV-C, Table III] The central claim is that monocular normal maps provide viewpoint-invariant geometric structure that improves matching. The chosen estimator (Omnidata DPT-Hybrid) is trained on indoor ScanNet/NYU and no evidence is provided that the predicted normals are accurate on outdoor drone/satellite imagery or consistent across the large viewpoint gap. The RGB+Normal vs. RGB comparison could also be explained by increased model capacity or low-level texture cues. Please add a quantitative evaluation of normal-map accuracy (e.g., against available 3D models or human-labeled normals) and a control experiment with a second non-geometric input (e.g., edge maps or a randomly initialized second branch) to isolate the geometric contribution.
  2. [§IV-B, Table II] The Baseline [8] rows are identical for Drone→Satellite and Satellite→Drone at every altitude (e.g., 55.65/61.92 at 150m). This is implausible because the two retrieval tasks have different query and gallery sets. Several AP values also appear inconsistent with the corresponding R@1 (e.g., LPN 150m Satellite→Drone: R@1=83.75 but AP=66.78). These issues compromise the reliability of the comparison table and must be corrected with properly recomputed numbers.
  3. [§IV-C, Fig. 6] The augmentation factor k is selected by measuring R@1/AP on University-1652 (k=4 is chosen because it is optimal in Fig. 6), and the same k is then used for the final results in Tables I and II. This is test-set tuning: there is no separate validation split. The reported SOTA numbers are therefore optimistically biased. Please use a train/validation/test protocol or report results across a range of k on a held-out split.
  4. [§III-C, Eq. (4)] Equation (4) contains unbalanced parentheses and apparent typos: the term 'f^m_r - p^m_n' inside SP(·) should presumably be f^m_r - f^m_n, and the use of ⊕ inside concat(·) is unclear. The equation is central to the DAFM module but cannot be parsed unambiguously. Please rewrite it with correct notation and define all operators.
  5. [§III-E vs. Contributions] The contribution list states that the 3D geographic augmentation technique works 'without the necessity of additional annotations,' yet Section III-E explicitly says 'we manually annotate the satellite image's trapezoidal imaging region r_s.' This is a manual annotation step. Please clarify whether the claim refers to not requiring new instance-level labels, and adjust the wording to avoid overstatement.
minor comments (6)
  1. [Table III] The text reports an average improvement of 1.7%, but the four improvements in Table III average about 2.7 percentage points. Please reconcile.
  2. [General] No error bars, multiple runs, or code are provided; the paper would benefit from reproducibility details (seeds, hardware, training epochs).
  3. [§III-D] The notation in the JCIA description (e.g., f'_q, F^{H×W→d}_s, dimension d=3) is difficult to follow. Please define explicitly.
  4. [Abstract] Typographical errors: 'UA V' should be 'UAV' and 'viewpoint ariations' should be 'viewpoint variations'.
  5. [Fig. 2] Figure 2 is dense and the font sizes in the sub-blocks are too small to read; please enlarge or vectorize.
  6. [Eq. (12)] The thresholds d_max and d_min are never specified. Please state the values used in the experiments.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity through test-set selection of the augmentation factor k; core RGB+normal fusion is an independent empirical contribution.

  1. fitted input called prediction [Section IV-C (Impact of Augmentation Factor k, Fig. 6); Tables I and II report Ours(k=4)]
    "From Fig. 6, it can be observed that as k increases from 1 to 4, both AP and R@1 continue to improve, reaching overall optimal performance when k=4."

    The augmentation factor k is selected by maximizing the exact R@1/AP metrics on University-1652, and the same selected value is then reported as 'Ours(k=4)' in the headline comparison tables. The reported augmentation gain is therefore not an independent prediction; it is the optimum of a curve used to pick the hyperparameter, making part of the final number fitted to the benchmark it is meant to validate.

full rationale

The paper is an empirical benchmark study, not a formal derivation, so most of its claims are supported by ablations and comparisons. The normal-image branch uses an external pretrained estimator (Omnidata/DPT-Hybrid), and the DAFM/JCIA modules are architectural designs with no built-in equivalence to the test labels. No load-bearing self-citation chain or imported uniqueness theorem is present. The only mild circularity is the selection of the augmentation factor k on the same dataset whose final metrics are then reported as state of the art; this is a fitted hyperparameter presented as a result, though it does not invalidate the independent RGB+normal ablation. Separately, the paper's contribution claim of 'without the necessity of additional annotations' is undercut by the manual satellite-region annotation described in Section III-E, but this is a consistency issue, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central result rests on four domain assumptions: transferability of ImageNet features, transferability of an indoor-trained normal estimator to aerial and satellite imagery, viewpoint-consistency of camera-frame normals, and COLMAP reconstruction quality. The method also has one tuned hyperparameter (k=4) selected on the target benchmark and several hand-set augmentation thresholds. No new physical entities are postulated.

free parameters (4)
  • p_s and p_c (DAFM learnable scale biases) = learned during training
    Equation (4) uses these to weight spatial and channel pooled statistics in the difference-aware fusion module.
  • augmentation factor k = 4
    Selected as optimal in Fig. 6 based on R@1 and AP on University-1652; used in all final reported results including SUES-200.
  • d_max and d_min thresholds and nine candidate centers = not specified
    Chosen by hand in Section III-E to limit cropping size in the augmentation; exact values are not reported.
  • spatial mapping dimension d in JCIA = 3
    Set to d=3 in experiments (Section III-D); no justification or ablation is provided.
assumptions (5)
  • domain assumption ImageNet-pretrained ConvNeXt-Base features transfer to drone and satellite imagery
    Section IV-A: the backbone is pretrained on ImageNet and used without domain adaptation analysis.
  • domain assumption Omnidata DPT-Hybrid normal estimation generalizes from indoor ScanNet/NYU Depth V2 to outdoor aerial and satellite RGB images
    Section III-B selects the estimator 'for its strong generalization', but no validation on the target aerial and satellite domains is given.
  • domain assumption Surface normals estimated independently in each camera frame are sufficiently viewpoint-invariant to serve as matching cues across the drone-satellite pose gap
    Sections I and III-B motivate normals as stable geometry, but no experiment isolates normal-map consistency under viewpoint change.
  • domain assumption COLMAP SfM recovers accurate 3D scene structure and camera poses from drone image sequences
    Section III-E relies on COLMAP reconstruction; no quality checks or failure cases are reported.
  • standard math Triplet loss plus cross-entropy loss and cosine similarity provide a valid retrieval objective
    Section III-F uses L_total = L_triplet + L_cross_entropy, a standard supervised metric-learning setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JRN-Geo: A Joint Perception Network based on RGB and Normal images for Cross-view Geo-localization." pith.science (2026). https://pith.science/paper/ZBYBXXVG

@misc{pith2026250905696,
  author       = {Pith},
  title        = {Pith review of: JRN-Geo: A Joint Perception Network based on RGB and Normal images for Cross-view Geo-localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZBYBXXVG}},
  note         = {Machine review of arXiv:2509.05696}
}
read the original abstract

Cross-view geo-localization plays a critical role in Unmanned Aerial Vehicle (UAV) localization and navigation. However, significant challenges arise from the drastic viewpoint differences and appearance variations between images. Existing methods predominantly rely on semantic features from RGB images, often neglecting the importance of spatial structural information in capturing viewpoint-invariant features. To address this issue, we incorporate geometric structural information from normal images and introduce a Joint perception network to integrate RGB and Normal images (JRN-Geo). Our approach utilizes a dual-branch feature extraction framework, leveraging a Difference-Aware Fusion Module (DAFM) and Joint-Constrained Interaction Aggregation (JCIA) strategy to enable deep fusion and joint-constrained semantic and structural information representation. Furthermore, we propose a 3D geographic augmentation technique to generate potential viewpoint variation samples, enhancing the network's ability to learn viewpoint-invariant features. Extensive experiments on the University-1652 and SUES-200 datasets validate the robustness of our method against complex viewpoint ariations, achieving state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2509.05696 by the authors.

Figure 1
Figure 1. A cross-view geo-localization method based on joint perception of RGB and normal images. The global descriptors are derived from RGB images captured by the camera and normal images generated through monocular normal estimation. Joint retrieval uses cosine similarity to measure the distance between drone and database descriptors. formation in complex scenes led to feature misassociation. To address this, researchers … view at source ↗
Figure 2
Figure 2. Overview of our proposed framework for cross-view geo-localization. (a) shows the pipeline of JRN-Geo, which integrates RGB and normal images for robust feature extraction. (b) DAFM focuses on the fusion of semantic and structural information. (c) JCIA strategy enhances spatial interaction and aggregates to obtain multi-feature representations. (d) The classifier module maps features for subsequent loss calculation.… view at source ↗
Figure 3
Figure 3. The 3D geographic augmentation technique. The left side shows the projection of camera imaging regions in 3D space, while the right side illustrates 2D homography transformation and cropping. map point c to the image pixel coordinates to obtain the mapped points ced , ces : ced = Hd, ces = Hsc, (11) where matrices Hd and Hs can be calculated from the corresponding vertex coordinates of the projection regions and the… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visualization of Augmentation Images. The left shows images before cropping and the right shows the augmented instance images with the augmentation factor k = 4. All augmented images are centered on the marked spatial points. Expansion Factor k (a) Expansion Factor k (…
Figure 4
Figure 4. Figure 4: Visualization of RGB and normal feature heatmaps. (a) shows the input Images. (b) shows that RGB focuses on the main building information, while normals specialize in capturing the global structure. (c) demonstrates the post-interaction feature maps integrating the spa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  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, pp. 1395– 1403, 2020

    work page 2020

  2. [1]

    Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization,

    S. Hu, M. Feng, R. M. Nguyen, and G. H. Lee, “Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7258–7267, 2018

    work page 2018

  3. [2]

    Dronet: Learning to fly by driving,

    A. Loquercio, A. I. Maqueda, C. R. Del-Blanco, and D. Scaramuzza, “Dronet: Learning to fly by driving,”IEEE Robotics and Automation Letters, vol. 3, no. 2, pp. 1088–1095, 2018

    2018

  4. [3]

    Lending orientation to neural networks for cross- view geo-localization,

    L. Liu and H. Li, “Lending orientation to neural networks for cross- view geo-localization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5624–5633, 2019

    work page 2019

  5. [4]

    Spatial-aware feature aggregation for image based cross-view geo-localization,

    Y . Shi, L. Liu, X. Yu, and H. Li, “Spatial-aware feature aggregation for image based cross-view geo-localization,”Advances in Neural Information Processing Systems, vol. 32, 2019

    work page 2019

  6. [5]

    Netvlad: Cnn architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,” inPro- ceedings of the IEEE conference on computer vision and pattern recognition, pp. 5297–5307, 2016

    work page 2016

  7. [6]

    Cross-view discriminative feature learning for person re-identification,

    A. Borgia, Y . Hua, E. Kodirov, and N. M. Robertson, “Cross-view discriminative feature learning for person re-identification,”IEEE Transactions on Image Processing, vol. 27, no. 11, pp. 5338–5349, 2018

    work page 2018

  8. [7]

    Learning feature descriptors using camera pose supervision,

    Q. Wang, X. Zhou, B. Hariharan, and N. Snavely, “Learning feature descriptors using camera pose supervision,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pp. 757–774, Springer, 2020

    work page 2020

Show all 34 references
  1. [9]

    A practical cross-view image matching method between uav and satellite for uav-based geo- localization,

    L. Ding, J. Zhou, L. Meng, and Z. Long, “A practical cross-view image matching method between uav and satellite for uav-based geo- localization,”Remote Sensing, vol. 13, no. 1, p. 47, 2020

    2020

  2. [10]

    Learning cross-view geo-localization embeddings via dynamic weighted decor- relation regularization,

    T. Wang, Z. Zheng, Z. Zhu, Y . Gao, Y . Yang, and C. Yan, “Learning cross-view geo-localization embeddings via dynamic weighted decor- relation regularization,”arXiv preprint arXiv:2211.05296, 2022

    2022 arXiv

  3. [11]

    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, 2021

    2021

  4. [12]

    A transformer-based feature segmentation and region alignment method for uav-view geo- localization,

    M. Dai, J. Hu, J. Zhuang, and E. Zheng, “A transformer-based feature segmentation and region alignment method for uav-view geo- localization,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 7, pp. 4376–4389, 2021

    2021

  5. [13]

    Joint representation learning and keypoint detection for cross-view geo-localization,

    J. Lin, Z. Zheng, Z. Zhong, Z. Luo, S. Li, Y . Yang, and N. Sebe, “Joint representation learning and keypoint detection for cross-view geo-localization,”IEEE Transactions on Image Processing, vol. 31, pp. 3780–3792, 2022

    2022

  6. [14]

    Mccg: A convnext- based multiple-classifier method for cross-view geo-localization,

    T. Shen, Y . Wei, L. Kang, S. Wan, and Y .-H. Yang, “Mccg: A convnext- based multiple-classifier method for cross-view geo-localization,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

    2023

  7. [15]

    Multi-level feedback joint representation learning network based on adaptive area elimination for cross-view geo-localization,

    F. Ge, Y . Zhang, L. Wang, W. Liu, Y . Liu, S. Coleman, and D. Kerr, “Multi-level feedback joint representation learning network based on adaptive area elimination for cross-view geo-localization,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  8. [16]

    Multibranch joint representation learning based on information fu- sion strategy for cross-view geo-localization,

    F. Ge, Y . Zhang, Y . Liu, G. Wang, S. Coleman, D. Kerr, and L. Wang, “Multibranch joint representation learning based on information fu- sion strategy for cross-view geo-localization,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–16, 2024

    2024

  9. [17]

    Surface normals in the wild,

    W. Chen, D. Xiang, and J. Deng, “Surface normals in the wild,” inProceedings of the IEEE International Conference on Computer Vision, pp. 1557–1566, 2017

    2017

  10. [18]

    Geonet: Geometric neural network for joint depth and surface normal estimation,

    X. Qi, R. Liao, Z. Liu, R. Urtasun, and J. Jia, “Geonet: Geometric neural network for joint depth and surface normal estimation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 283–291, 2018

    2018

  11. [19]

    Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,

    A. Eftekhar, A. Sax, J. Malik, and A. Zamir, “Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10786–10796, 2021

    2021

  12. [20]

    Monosdf: Exploring monocular geometric cues for neural implicit surface recon- struction,

    Z. Yu, S. Peng, M. Niemeyer, T. Sattler, and A. Geiger, “Monosdf: Exploring monocular geometric cues for neural implicit surface recon- struction,”Advances in neural information processing systems, vol. 35, pp. 25018–25032, 2022

    2022

  13. [21]

    Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors,

    S. Hwang, M.-J. Kim, T. Kang, J. Kang, and J. Choo, “Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors,”arXiv preprint arXiv:2407.02945, 2024

    2024 arXiv

  14. [22]

    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, 2023

    2023

  15. [23]

    Structure-from-motion revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-from-motion revisited,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 4104–4113, 2016

    2016

  16. [24]

    Uav-satellite view syn- thesis for cross-view geo-localization,

    X. Tian, J. Shao, D. Ouyang, and H. T. Shen, “Uav-satellite view syn- thesis for cross-view geo-localization,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 7, pp. 4804–4815, 2021

    2021

  17. [25]

    Sample4geo: Hard negative sampling for cross-view geo-localisation,

    F. Deuser, K. Habel, and N. Oswald, “Sample4geo: Hard negative sampling for cross-view geo-localisation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16847– 16856, 2023

    2023

  18. [26]

    Deep- mvs: Learning multi-view stereopsis,

    P.-H. Huang, K. Matzen, J. Kopf, N. Ahuja, and J.-B. Huang, “Deep- mvs: Learning multi-view stereopsis,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 2821– 2830, 2018

    2018

  19. [27]

    En- hancing visual place recognition with multi-modal features and time- constrained graph attention aggregation,

    Z. Wang, Y . Zhang, X. Zhao, J. Ning, D. Zou, and M. Pei, “En- hancing visual place recognition with multi-modal features and time- constrained graph attention aggregation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 15914–15921, IEEE, 2024

    2024

  20. [28]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 12179–12188, 2021

    2021

  21. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016

    2016

  22. [30]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  23. [31]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 5828–5839, 2017

    2017

  24. [32]

    Indoor seg- mentation and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor seg- mentation and support inference from rgbd images,” inComputer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pp. 746– 760, Springer, 2012

    2012

  25. [33]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11976–11986, 2022

    2022

  26. [34]

    Uav’s status is worth considering: A fusion representations matching method for geo- localization,

    R. Zhu, M. Yang, L. Yin, F. Wu, and Y . Yang, “Uav’s status is worth considering: A fusion representations matching method for geo- localization,”Sensors, vol. 23, no. 2, p. 720, 2023

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.