Pith. sign in

REVIEW 2 major objections 6 minor 41 references

Explicitly conditioning query embeddings on trajectory geometry—waypoint headings and origin–destination bearing—is the key to accurate cross-view geo-localization from both videos and route descriptions, not simply larger neural backbones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:10 UTC pith:6QFKONRR

load-bearing objection A genuinely new benchmark and a solid text-retrieval method, but the video SOTA claim outruns the evidence — the dataset and text-side experiments carry the paper. the 2 major comments →

arxiv 2607.15491 v1 pith:6QFKONRR submitted 2026-07-16 cs.CV

Trajectory-aware Cross-view Geo-localization with Sequential Observations

Tianyi Gao , Jiayu Lin , Danielle Beaulieu , Nathan Jacobs This is my paper
classification cs.CV
keywords cross-view geo-localizationtrajectory-aware retrievalvideo geo-localizationtext geo-localizationsatellite imagerycontrastive learningtrajectory modulationmultimodal benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that cross-view geo-localization—matching ground-level observations to satellite imagery—can be unified for two complementary sequential inputs: video clips and natural-language route descriptions, and that these modalities reinforce each other when trained together. The central new idea is TrajMod, a small module that converts per-waypoint headings and the overall start-to-end bearing into a geometry embedding, then uses it to rescale and shift the query embedding before retrieval. On a new benchmark of about 39,000 video–text–satellite triplets, the method achieves state-of-the-art recall on both tasks. Ablations attribute the large gains specifically to trajectory-conditioned modulation, not backbone choice or prompt engineering.

Core claim

The central claim is that explicit trajectory geometry—available from IMUs and compasses on modern devices but usually stripped from web videos—should be injected into cross-view retrieval, and that a lightweight FiLM-style modulation is the effective way to do it. On the SeqGeo-VL benchmark, TrajLoc achieves R@1 of 12.09% on video geo-localization and 2.52% on text geo-localization, outperforming prior methods by a wide margin. The paper further shows that TrajMod lifts text R@1% from 26.15% to 45.48% (+19.33 points), whereas adding trajectory details as text prompts yields only +1.34 points, indicating that geometric modulation is fundamentally more effective than language instructions for

What carries the argument

TrajMod (trajectory-conditioned modulation) is the central mechanism. It builds a trajectory geometry embedding by applying Fourier features to waypoint heading angles and the origin–destination bearing, then passes it through a modality-specific MLP that predicts per-dimension scale and shift parameters. These parameters modulate the video or text query embedding via element-wise multiplication and addition (FiLM-style). Because both modalities are conditioned on the same trajectory geometry, TrajMod acts as a shared spatial anchor, enabling mutual reinforcement between visual and textual sequential observations during co-training.

Load-bearing premise

The method assumes trajectory geometry—waypoint headings and origin–destination bearing—is available at test time as an input modality, and that without it TrajMod cannot be applied, so the reported gains may not transfer to users who only provide a text description with no sensor metadata.

What would settle it

Run the same retrieval pipeline on SeqGeo-VL with the trajectory geometry inputs corrupted (e.g., random headings or swapped bearings) and compare recall to the intact TrajMod. If R@1 does not drop measurably, then the trajectory-conditioned modulation is not the source of the reported gains and the central claim is refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single unified model can serve both video-based and text-based geo-localization, avoiding the need for separate task-specific systems.
  • Route descriptions become a practical fallback modality when video is unavailable, sparse, or bandwidth-limited, since text-only retrieval benefits from the same trajectory conditioning.
  • Explicit geometric priors from IMUs and compasses should be treated as a standard input modality for vision-language geo-localization, not an optional extra.
  • Co-training video and text encoders against a shared satellite encoder with a two-stage curriculum yields mutual gains that exceed training either modality alone.
  • The large gap between TrajMod and prompt-based trajectory injection suggests that geometric modulation is a more effective interface than natural language for injecting spatial layout into embeddings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If trajectory geometry is the true driver, then randomized or shuffled headings and bearings should substantially degrade retrieval accuracy—a falsifiable prediction that could be tested in a follow-up experiment.
  • Because the paper shows gains even with a single frame but larger gains with longer sequences, the modulation mechanism may be especially valuable in long, visually aliased trajectories where appearance alone is ambiguous.
  • The same modulation scheme could transfer to other cross-view tasks (e.g., drone-to-satellite matching or indoor localization) wherever heading and bearing are measurable, suggesting a broader design pattern for spatially grounded retrieval.
  • A natural extension would be to predict trajectory geometry from the video or text itself, removing the test-time sensor requirement and extending the method to settings where IMU or compass data is not available.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. This paper introduces SeqGeo-VL, a benchmark of 38,863 video-text-satellite triplets built from SeqGeo by adding route-level descriptions produced with a Qwen-VLM/LLM annotation pipeline and a 1% human quality check. It also proposes TrajLoc, a CLIP-initialized three-encoder framework using a two-stage curriculum (video-satellite pretraining, then text-satellite alignment with a cosine-similarity regularizer), and TrajMod, a FiLM-style module that modulates video/text query embeddings with Fourier-encoded waypoint headings and origin-destination bearing. Experiments on SeqGeo-VL report video R@1 of 12.09% (vs SeqGeo† 8.14%) and text R@1 of 2.52% (vs CrossText2Loc 0.98%), with ablations attributing a large text R@1% gain (+19.33) to TrajMod.

Significance. If the results hold, SeqGeo-VL is a useful resource for route-description geo-localization, and TrajMod is a clean, lightweight mechanism for injecting trajectory geometry into CLIP-style embeddings. The authors commit to releasing code, model weights, and the dataset, and the Table 3 text baselines appear to be fully fine-tuned on SeqGeo-VL, making that comparison internally fair. The ablations are consistent and clearly identify trajectory geometry as the main driver of the reported gains. However, the video state-of-the-art claim is not supported by the evidence as presented, and the text SOTA results depend on a test-time trajectory-geometry assumption that should be stated prominently. With those issues addressed, the contribution would be significant for embodied cross-view geo-localization.

major comments (2)
  1. [Table 2 / Sec. 5.2] The video SOTA claim is based on cross-dataset numbers. The caption states that, apart from SeqGeo† and Qwen3-VL-Embedding*, results are 'excerpted from the original papers.' GARet and FlexGeo were evaluated on GAMa/BDD100K and Set-CVGL respectively, not on SeqGeo-VL. Their R@1 values (3.34%, 3.51%) are therefore not comparable with TrajLoc's 12.09% measured on SeqGeo-VL. The only same-benchmark video comparison is against SeqGeo† (8.14%), which is an older method reimplemented by the authors. Please evaluate GARet and FlexGeo on the SeqGeo-VL test split with the same protocol, or revise the abstract and conclusion to claim only a comparison with SeqGeo† and Qwen3-VL-Embedding on SeqGeo-VL.
  2. [Sec. 4.2, Eq. (4) and Sec. 5.3] The text SOTA results assume trajectory geometry is available at test time. TrajMod consumes z_traj (waypoint headings and OD bearing) and modulates the text embedding via Eq. (4); a bare route description does not provide these signals. The paper's framing that route descriptions are 'often the only input available' is in tension with this, and Table 4 shows text R@1 drops from 2.52% to 0.98% when TrajMod is removed. Please either report text results without trajectory geometry in the main comparison or explicitly state that the text SOTA result assumes IMU/compass metadata, and discuss the resulting application scope.
minor comments (6)
  1. [Table 2 caption] The caption should specify the dataset and evaluation protocol for each excerpted method; as written, a reader can easily misread GARet/FlexGeo numbers as SeqGeo-VL results.
  2. [Sec. 5.4 / Fig. 6] Fig. 6 shows frame counts from 1 to 6, but the later 'Co-sequenced Multimodal Query' paragraph refers to 7 frames. Make the experimental protocols consistent or clarify that the two experiments use different maximum sequence lengths.
  3. [Eq. (3)] Define the Fourier feature dimension K and state whether angles are in radians; also write the factors unambiguously, e.g., sin(2^{k-1} * pi * alpha).
  4. [Table 1 / Sec. 3.2] SeqGeo-VL is described as having 38,863 triplets, but Table 1 lists Qry=118K. Clarify the relation between triplets and the 118K video/text query instances so the reader is not confused.
  5. [Sec. 3.2] Gwet's AC2 is mentioned for inter-annotator agreement, but no AC2 value is reported. Reporting the actual coefficient would strengthen the quality-control claim.
  6. [Sec. 5.1] The sentence 'consistent except for a few noted exceptions' is vague; list the exceptions or remove it.

Circularity Check

0 steps flagged

No circularity: TrajLoc's gains come from trajectory-conditioned modulation and co-training on an independent train/test split; remaining caveats are external-validity concerns, not definitional circularity.

full rationale

The paper's derivation chain is not circular. TrajMod (Eq. 4) takes trajectory geometry ztraj as an input feature derived from waypoint headings and OD bearing, and learns modulation parameters gamma, beta via an MLP. ztraj is a raw sensor-derived input, not a fitted quantity produced from the retrieval labels, so calling the resulting retrieval a 'prediction' is legitimate. The contrastive losses (Eq. 5) are trained on a train/test split of SeqGeo-VL, and retrieval is evaluated against independent satellite gallery items; there is no equation-level equivalence between an input and an output. The route descriptions are generated from ground frames and trajectory metadata, not from the satellite labels, so the text-to-satellite task is not self-definitional. The paper does contain some self-citations (e.g., refs. [24], [28], [29] involve Nathan Jacobs), but these are dataset or related-work citations and are not load-bearing for the central claim. The main caveats—that the benchmark is author-constructed and that Table 2 mixes results from different benchmarks for GARet and FlexGeo—are concerns about external validity and evidence quality, not about circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper does not introduce new physical entities. It relies on the SeqGeo dataset, CLIP backbones, generated text annotations, and the stated availability of trajectory sensors. The main free parameters are standard training hyperparameters and the blend weight; the key domain assumption is that trajectory geometry is available at inference.

free parameters (3)
  • InfoNCE temperature
    Temperature in the symmetric contrastive loss (Eq. 5); value not reported.
  • Fourier feature dimension K
    Number of frequency bands in angle encoding (Eq. 3); not specified.
  • Blend weight alpha = 0.4
    Weight for linearly blending video and text query embeddings in co-sequenced query experiments (Sec. 5.4); chosen by hand.
axioms (5)
  • domain assumption SeqGeo provides accurate trajectory-aligned video and satellite pairs.
    SeqGeo-VL is built on SeqGeo; if the underlying geo-alignment is wrong, all localization results are invalid (Sec. 3.2).
  • domain assumption Trajectory geometry (headings and OD bearing) is available at test time from IMU/compass.
    TrajMod requires z_traj; explicitly assumed in Sec. 1 and Sec. 4.2.
  • domain assumption VLM/LLM-generated route descriptions faithfully represent the trajectories.
    Texts are generated by Qwen3-VL and Qwen3-30B; only 400 of 38,863 triplets are human-verified (Sec. 3.2).
  • domain assumption CLIP ViT-L/14 provides a reasonable shared embedding space for ground and satellite views.
    All three encoders are initialized from CLIP (Sec. 4.1).
  • domain assumption Satellite imagery at zoom level 20 contains enough spatial detail to distinguish routes.
    Gallery patches are at zoom level 20 (Sec. 3.2).

pith-pipeline@v1.3.0-alltime-deepseek · 12088 in / 12578 out tokens · 120771 ms · 2026-08-01T23:10:21.203763+00:00 · methodology

0 comments
read the original abstract

Cross-view geo-localization matches ground-level observations against geo-tagged satellite imagery. Recent methods show that sequential queries such as video clips yield richer spatiotemporal cues than single images, yet they overlook a complementary sequential modality: route descriptions -- which capture the same trajectory at a higher level of abstraction and are often the only input available (e.g., a user directing an autonomous vehicle to a pickup point). To bridge this gap, we introduce SeqGeo-VL, a dataset of $\sim$39K video-text-satellite triplets, and TrajLoc, a unified framework capable of processing both video clips and route descriptions. By leveraging both dense visual and abstract linguistic semantics, TrajLoc enables these modalities to mutually reinforce cross-view matching. We further propose TrajMod, a lightweight module that conditions query embeddings on trajectory geometry, yielding spatially-aware representations. Experiments show that TrajLoc achieves substantial gains over state-of-the-art methods on both video and text geo-localization. The project page is available at https://humblegamer.github.io/trajloc/.

Figures

Figures reproduced from arXiv: 2607.15491 by Danielle Beaulieu, Jiayu Lin, Nathan Jacobs, Tianyi Gao.

Figure 1
Figure 1. Figure 1: Motivation for our approach, which extends existing cross-view geo-localization methods to sequential route descriptions and explicitly incorporates trajectory infor￾mation into video/text queries, producing spatially-grounded representations. However, existing cross-view geo-localization datasets lack route descriptions, let alone video–text–satellite triplets. This restricts flexibility and leaves the po… view at source ↗
Figure 2
Figure 2. Figure 2: Word cloud and distribution of the extracted keywords within the SeqGeo-VL dataset [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the TrajLoc framework. Left: Encoder adaptation with a two-stage curriculum. Three CLIP-initialized encoders produce video (fv), text (ft), and satellite (fs) embeddings. Stage 1 trains the video and satellite encoders with a contrastive loss Lctr; Stage 2 freezes the video encoder and aligns the text encoder, with a regularization loss Lreg to prevent satellite-encoder drift. Right: All three … view at source ↗
Figure 5
Figure 5. Figure 5: Cross-view geo-localization with and without TrajMod. Blue highlights the spatial cues, while bold denotes the observed objects. As observed in Top-3 candidates, TrajMod enables more layout-aligned retrieval results. noise that is not shared across views and encourages both f ′ v and f ′ t to preserve the common, geometry-grounded factors injected by ztraj. 5 Experiments We evaluate TrajLoc on the SeqGeo-V… view at source ↗
Figure 6
Figure 6. Figure 6: Video geo-localization performance with varying frame number and TrajMod. Location Prior Context. In practical applications, users or mobile robots often op￾erate with a location prior, such as knowing their rough geographic district when GPS signals are severely obscured. To simulate these conditions, we evaluate re￾trieval performance under varying gallery sizes, effectively narrowing the search space ba… view at source ↗
Figure 7
Figure 7. Figure 7: Performance with different gallery sizes given a location prior. Synergy between Co-training and TrajMod. We investigate the mutual benefits of jointly training cross-view geo-localization with visual or linguistic sequential observations, and how TrajMod amplifies this synergy. As reported in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 6 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674–3683 (2018) 16 T. Gao et al

    2018

  2. [2]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    arXiv preprint arXiv:2504.13181 (2025)

    Bolya, D., Huang, P.Y., Sun, P., Cho, J.H., Madotto, A., Wei, C., Ma, T., Zhi, J., Rajasegaran, J., Rasheed, H., et al.: Perception encoder: The best visual em- beddings are not at the output of the network. arXiv preprint arXiv:2504.13181 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    Advances in Neural Information Processing Systems37, 135062–135093 (2024)

    Cheng, A.C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., Liu, S.: Spatialrgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems37, 135062–135093 (2024)

    2024

  5. [5]

    In: Euro- pean Conference on Computer Vision

    Chu, M., Zheng, Z., Ji, W., Wang, T., Chua, T.S.: Towards natural language- guided drones: Geotext-1652 benchmark with spatial relation matching. In: Euro- pean Conference on Computer Vision. pp. 213–231. Springer (2024)

    2024

  6. [6]

    Image and Vision Computing149, 105171 (2024)

    Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. Image and Vision Computing149, 105171 (2024)

    2024

  7. [7]

    The international journal of robotics research32(11), 1231–1237 (2013)

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The international journal of robotics research32(11), 1231–1237 (2013)

    2013

  8. [8]

    Iclr1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. Iclr1(2), 3 (2022)

    2022

  9. [9]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Kolmet, M., Zhou, Q., Ošep, A., Leal-Taixé, L.: Text2pos: Text-to-point-cloud cross-modal localization. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 6687–6696 (2022)

    2022

  10. [10]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Lin, T.Y., Belongie, S., Hays, J.: Cross-view image geolocalization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 891–898 (2013)

    2013

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

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

    2019

  12. [12]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Liu, M., He, H., Ricci, E., Wu, W., Zhou, B.: Urbanverse: Scaling urban simula- tion by watching city-tour videos. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  13. [13]

    arXiv preprint arXiv:2104.08860 (2021)

    Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860 (2021)

    Pith/arXiv arXiv 2021

  14. [14]

    arXiv preprint arXiv:1807.03748 (2018)

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  15. [15]

    In: Proceedings of the AAAI conference on artificial intelligence

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual rea- soning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)

    2018

  16. [16]

    In: European Conference on Computer Vision

    Pillai, M.S., Rizve, M.N., Shah, M.: Garet: cross-view video geolocalization with adapters and auto-regressive transformers. In: European Conference on Computer Vision. pp. 466–483. Springer (2024)

    2024

  17. [17]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Sastry, S., Choudhury, A., Madala, P., Su, G.M.: Probabilistic multimodal learning with bayesian disagreements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 7546–7555 (June 2026) Trajectory-aware Cross-view Geo-localization 17

    2026

  19. [19]

    In: Asian Conference on Computer Vision

    Shi, Y., Yu, X., Wang, S., Li, H.: Cvlnet: Cross-view semantic correspondence learning for video-based camera localization. In: Asian Conference on Computer Vision. pp. 123–141. Springer (2022)

    2022

  20. [20]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Tian, H., Meng, J., Zheng, W.S., Li, Y.M., Yan, J., Zhang, Y.: Loc4plan: Locating before planning for outdoor vision and language navigation. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 4073–4081 (2024)

    2024

  21. [21]

    In: Euro- pean Conference on Computer Vision

    Vyas, S., Chen, C., Shah, M.: Gama: Cross-view video geo-localization. In: Euro- pean Conference on Computer Vision. pp. 440–456. Springer (2022)

    2022

  22. [22]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Z., Zhou, S., He, S., Huang, H., Yang, L., Zhang, Z., Cheng, X., Ji, S., Jin, T., Zhao, H., et al.: Spatialclip: Learning 3d-aware image representations from spatially discriminative language. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 29656–29666 (2025)

    2025

  23. [23]

    In: Forty-third International Conference on Machine Learn- ing (2026),https://openreview.net/forum?id=BkZ40FNKlL

    Wei,W.,Gui,Z.,Peng,D.,Ye,T.,Wu,H.:Variationaladapterforcross-modalsim- ilarity representation. In: Forty-third International Conference on Machine Learn- ing (2026),https://openreview.net/forum?id=BkZ40FNKlL

    2026

  24. [24]

    In: Proceedings of the IEEE International Conference on Com- puter Vision

    Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocalization with aerial reference imagery. In: Proceedings of the IEEE International Conference on Com- puter Vision. pp. 3961–3969 (2015)

    2015

  25. [25]

    ISPRS Journal of Photogrammetry and Remote Sensing233, 328–345 (2026).https://doi.org/https://doi.org/10.1016/j

    Wu, Q., Xia, P., Yu, L., Liu, Y., Xiong, M., Zhong, L., Chen, J., Yang, M., Zhang, Y., Wan, Y.: Set-cvgl: A new perspective on cross-view geo-localization with un- ordered ground-view image sets. ISPRS Journal of Photogrammetry and Remote Sensing233, 328–345 (2026).https://doi.org/https://doi.org/10.1016/j. isprsjprs.2026.01.037,https://www.sciencedirect....

    doi:10.1016/j 2026

  26. [26]

    ISPRS Journal of Photogrammetry and Remote Sensing227, 1–11 (2025)

    Xia, P., Yu, L., Wan, Y., Wu, Q., Chen, P., Zhong, L., Yao, Y., Wei, D., Liu, X., Ru, L., et al.: Cross-view geo-localization with panoramic street-view and vhr satellite imagery in decentrality settings. ISPRS Journal of Photogrammetry and Remote Sensing227, 1–11 (2025)

    2025

  27. [27]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xia, Y., Shi, L., Ding, Z., Henriques, J.F., Cremers, D.: Text2loc: 3d point cloud localization from natural language. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14958–14967 (2024)

    2024

  28. [28]

    In: Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N

    Xing, E., Stylianou, A., Pless, R., Jacobs, N.: Quari: Query adaptive re- trieval improvement. In: Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N. (eds.) Advances in Neural Informa- tion Processing Systems. vol. 38, pp. 62890–62911. Curran Associates, Inc. (2025),https : / / proceedings . neurips . cc / paper _ files / p...

    2025

  29. [29]

    Transactions on Machine Learning Research

    Xiong, Z., Xing, X., Workman, S., Khanal, S., Jacobs, N.: Mixed-view panorama synthesis using geospatially guided diffusion. Transactions on Machine Learning Research

  30. [30]

    arXiv preprint arXiv:2505.09388 (2025)

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    Yang, J., Yang, S., Gupta, A.W., Han, R., Fei-Fei, L., Xie, S.: Thinking in space: How multimodal large language models see, remember, and recall spaces. In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 10632– 10643 (2025)

    2025

  32. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Ye, J., Lin, H., Ou, L., Chen, D., Wang, Z., Zhu, Q., He, C., Li, W.: Where am i? cross-view geo-localization with natural language descriptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5890–5900 (2025) 18 T. Gao et al

    2025

  33. [33]

    In: European Conference on Computer Vision

    Ye, J., Lv, Z., Li, W., Yu, J., Yang, H., Zhong, H., He, C.: Cross-view image geo- localization with panorama-bev co-retrieval network. In: European Conference on Computer Vision. pp. 74–90. Springer (2024)

    2024

  34. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 2636–2645 (2020)

    2020

  35. [35]

    Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936 (2022)

    Pith/arXiv arXiv 2022

  36. [36]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language im- age pre-training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11975–11986 (2023)

    2023

  37. [37]

    arXiv preprint arXiv:2602.08342 (2026)

    Zhang, J., Yu, X., Fang, Y., Stouffs, R., Trivic, Z.: Urbangraphembeddings: Learn- ing and evaluating spatially grounded multimodal embeddings for urban science. arXiv preprint arXiv:2602.08342 (2026)

    arXiv 2026

  38. [38]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Zhang, X., Sultani, W., Wshah, S.: Cross-view image sequence geo-localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2914–2923 (2023)

    2023

  39. [39]

    In: Proceedings of the 28th ACM interna- tional conference on Multimedia

    Zheng, Z., Wei, Y., Yang, Y.: University-1652: A multi-view multi-source bench- mark for drone-based geo-localization. In: Proceedings of the 28th ACM interna- tional conference on Multimedia. pp. 1395–1403 (2020)

    2020

  40. [40]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhu, S., Shah, M., Chen, C.: Transgeo: Transformer is all you need for cross-view image geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1162–1171 (2022)

    2022

  41. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhu, S., Yang, T., Chen, C.: Vigor: Cross-view image geo-localization beyond one- to-one retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3640–3649 (2021)

    2021