Pith. sign in

REVIEW 4 major objections 5 minor 68 references

GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space

T0 review · 4 major / 5 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 single shared embedding for images, time, and location lets GT-Loc predict hour, month, and GPS from pixels alone, outperforming time predictors that are given the true location at inference.

desk verdict A genuinely new retrieval-based joint time-and-location method with a sensible toroidal metric-learning loss, but the headline claim over prior methods rests on reimplemented baselines whose protocols are not fully available. read the letter →

arxiv 2507.10473 v2 pith:X2KZU5YY submitted 2025-07-14 cs.CV

David G. Shatwell , Ishan Rajendrakumar Dave , Sirnam Swetha , Mubarak Shah This is my paper · ORCID
classification cs.CV
keywords timestamppredictiontime-of-capturegeo-localizationjointembeddingtemporalmetriclearningtoroidaldistanceimageretrievalzero-shotevaluation
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

GT-Loc tries to establish that capture time and geographic location can be predicted together from a single image, without any metadata, by treating both as retrieval problems in one shared embedding space. The paper claims its joint model reaches a mean month error of 1.40 and hour error of 2.72 hours on unseen cameras, and that this beats prior time-prediction methods that are given the ground-truth GPS coordinates during inference. The reason this matters is that real applications—forensics, metadata repair, and retrieval—often lack trustworthy timestamps or locations, and current methods depend on exactly the metadata they are meant to verify. The central move is a new Temporal Metric Learning loss that replaces hard positive-negative pairs with soft similarity targets shaped by the cyclic distance between times on a torus.

What carries the argument

The load-bearing object is the Temporal Metric Learning (TML) loss. It converts each batch's image-time cosine-similarity row into a softmax distribution over the batch's time embeddings, and the corresponding target distribution is built from toroidal distances $\delta_{i,j}=\sqrt{\sum_{\alpha\in\{\theta,\phi\}} \min(1-|\Delta\alpha_{i,j}|,|\Delta\alpha_{i,j}|)^2}$, normalized so that nearby times receive higher target probability; the loss is the cross-entropy between these two distributions. The toroidal distance is what makes the supervision respect the cyclical nature of hours and months, avoiding the overestimation of plain Euclidean distance between, say, 23:00 and 01:00. Around this sits the architecture: a frozen CLIP ViT-L/14 image encoder with a trainable MLP, a location encoder using Equal Earth Projection plus random Fourier features, and a structurally identical time encoder fed with normalized cyclic month-hour pairs.

What would settle it

Take a held-out set of image pairs with known capture times and compute the rank correlation between the frozen image encoder's cosine similarity and the toroidal time difference before any training. If the correlation is near zero, the smoothness premise TML relies on is absent, so the soft-target loss would have no valid signal to learn from.

Watch

Extended reading notes

Core claim

The central claim is that aligning images, times, and locations in a common embedding space solves both tasks at once, and that time alignment is best learned not by contrastive learning but by a metric loss whose targets are soft: each image is pushed toward all times in the batch, with force proportional to how close those times are, measured cyclically over a toroidal manifold. Because months and hours are periodic, the paper represents each timestamp as a normalized month-hour pair and encodes it with random Fourier features, mirroring the location encoder. On the SkyFinder unseen-camera benchmark, GT-Loc reports 1.40 month error, 2.72 hour error, and a Time Prediction Score of 77.00, surpassing replicated baselines including those that take GPS as input, and it also reaches 14.41 percent top-1 accuracy at 1 km on Im2GPS3k, competitive with specialist geo-localizers. The paper further claims the shared space enables composed query retrieval (time plus location to image) and text-based time and location prediction using the frozen text encoder.

Load-bearing premise

The load-bearing premise is that visual appearance changes smoothly and continuously with time, so that using similarity proportional to toroidal time difference as a training target is valid rather than a distortion of the embedding.

Editorial extensions

If this is right

  • Time prediction no longer needs GPS at inference: GT-Loc's image-only retrieval beats the three replicated baselines, all of which consume geo-location as input.
  • Joint training with location helps time: the full GT-Loc outperforms its own time-only variant (77.00 vs 75.49 TPS), indicating cross-modal supervision enriches temporal embeddings.
  • The retrieval formulation has bounded cost: a 4,000-sample time gallery already gives strong performance, and a 500,000-sample gallery adds only about 950 MB and roughly 0.5 TFLOP of retrieval overhead.
  • Degradation under scarce and noisy data is graceful: with 5% of training data the TPS drops from 77.00 to 72.07, and with label noise at $\sigma=2$ it remains at 73.81.
  • The same space supports composed and text queries: averaging time and location embeddings retrieves matching images at 6.69% R@1 on unseen SkyFinder cameras, and CLIP text queries produce sensible time and location distributions for phrases like "warm summer evening."

Reading between the lines

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

  • A direct extension would be to apply TML to other cyclic targets—day of week, season, tide phase, or traffic state—wherever visual appearance varies smoothly; the same soft-target plus toroidal-distance recipe should transfer.
  • Because the time encoder shares the location encoder's architecture, the framework could be extended to additional continuous metadata such as altitude, weather, or camera heading by adding encoders and the same metric objective.
  • The comparison suggests a testable scaling claim: if CLIP embeddings are already smooth in time, then TML should require very little data to beat hard-negative losses; the 5%-data row (72.07 TPS) is consistent with that, but the paper does not isolate this mechanism.
  • A falsifiable practical prediction is that TML's advantage over contrastive losses grows as the time gallery becomes denser, since soft targets can exploit intermediate neighbors that hard positive and negative assignment would discard.
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

4 major / 5 minor

Summary. GT-Loc proposes a retrieval-based framework that jointly predicts capture time (month and hour) and geo-location from a single image. The method uses CLIP for image features, an RFF-based location encoder, and an RFF-based time encoder operating on a normalized month-hour representation on a toroidal manifold. The key technical novelty is Temporal Metric Learning (TML), which replaces hard positive/negative contrastive pairs with soft targets derived from cyclic time differences and optimizes a cross-entropy between normalized image-time similarities and normalized time-difference targets. The paper reports zero-shot time prediction results on unseen SkyFinder cameras (Table 1), geo-localization results on Im2GPS3k and GWS15k (Table 2), ablations over loss functions (Table 3), robustness studies (Tables 4 and 5), and a compositional retrieval experiment (Table 6), plus qualitative text-based retrieval examples.

Significance. The TML idea is simple, well-motivated, and potentially useful: treating month and hour as a toroidal manifold and using smooth soft targets rather than hard negatives is a sensible way to handle temporally adjacent visual similarity. The ablation in Table 3 provides initial evidence that the cyclic TML loss outperforms a Euclidean variant and several contrastive baselines. The robustness and compositional retrieval experiments add breadth. The paper does not ship code or released splits, and the headline comparison to prior time-prediction methods rests on reimplementations whose original protocols are partly unavailable. If the authors correct the TML formula, release or specify the baseline protocols precisely, and add variance estimates, the contribution could be solid and of interest to the retrieval and geo-localization community.

major comments (4)
  1. [Section 4.1, Table 1, Supplementary Section 21] The headline claim that GT-Loc surpasses previous time prediction methods, even those using ground-truth geo-location, is not established by Table 1 because all three baselines are author reimplementations. Supplementary Section 21 states that Zhai et al. provide no code or exact splits, Salem et al. omit split details and code, and Padilha et al. do not provide the cross-camera split; the "w/ CLIP" variants also replace the original InceptionV2/DenseNet-121 backbones with frozen CLIP ViT-L/14. These differences in training data, split, and backbone mean the comparison is not controlled. The authors should either evaluate against original checkpoints/protocols where available, release their exact splits and reimplementations, or explicitly scope the headline claim to the reimplemented baselines.
  2. [Section 3.3, Eq. (8)] Equation (8) as printed is not a valid probability distribution. If q_i[j] is taken literally as 1 minus exp(delta_{i,j}) divided by a sum of exp(delta_{i,k}), the numerator is negative for every positive delta and the denominator does not match the numerator, so the cross-entropy in Eq. (9) is undefined. The intended soft-target formula must be stated correctly, for example q_i[j] = exp(-delta_{i,j}) / sum_k exp(-delta_{i,k}), or an explicitly renormalized variant of 1 - softmax(delta). This is the core of the proposed TML loss and must be corrected.
  3. [Tables 1-6, Section 4] All experiments are reported as single runs without error bars, standard deviations, or significance tests. Several differences that support the paper's claims are small in absolute terms, e.g., TPS 77.00 vs 75.49 for TimeLoc in Table 1 and TPS 77.00 vs 76.61 for the daily ToY scale in Table 9. Without multiple seeds or variance estimates it is unclear whether these improvements are reliable. The authors should report mean +/- standard deviation over at least three seeds and, where feasible, a paired significance test, especially for the baseline comparisons in Table 1.
  4. [Supplementary Section 12] The training set is filtered to remove night-time and indoor images, but the paper does not state whether the zero-shot SkyFinder test subset is filtered in the same way. If the test set is also day/outdoor only, the reported time-prediction accuracy covers a narrower setting than the Introduction's general claims about timestamp prediction. The authors should specify the test-time filtering procedure and discuss how the removal of night and indoor images affects the scope of the conclusions.
minor comments (5)
  1. [Section 4.4] The sentence "Table 1 shows quantitative results for these baselines" should refer to Table 6, which is the table reporting compositional retrieval results.
  2. [Section 3.1, Eq. (1)] The index i is used both for the sample index and for the RFF scale sigma_i in Eq. (1); using a different symbol for the scale, such as m, would avoid confusion.
  3. [Section 3.2, Eqs. (3)-(4)] The convention for D(m_i) in leap years is unspecified; the authors should state whether D(m_i) is fixed to a non-leap year or computed per year.
  4. [Supplementary Section 16, Table 10] The claim that geo-localization errors tend to be higher in lower-population-density areas is not monotonic on GWS15k: the densest group (>=10k) has median error 1035.81 km, higher than the two middle groups. The text should be revised to reflect this non-monotonic pattern.
  5. [Eq. (11)] The notation eE_{ToY} and eE_{ToD} in the TPS equation is awkward; using hat{e} or a clearer symbol for the normalized errors would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the TML loss is a supervised objective, and GT-Loc's time-prediction claims are supported by ablations and external benchmarks, not by a self-referential derivation.

full rationale

The central time-prediction mechanism is not circular. The TML objective (Eqs. 6-9) derives soft targets q_i from ground-truth timestamps via the toroidal distance δ_i,j and trains image embeddings to match those targets; inference is a nearest-neighbor retrieval against a gallery of time embeddings, so the reported month/hour errors and TPS are empirical outcomes of learned embeddings, not fitted values renamed as predictions. The TPS metric (Eq. 11) uses the same cyclic-time notion as the loss, but the paper also reports plain month and hour errors, and using a task-aligned metric is standard for regression-style evaluation. The location branch is adopted from GeoCLIP [37], a same-group prior work, but this is an explicit architectural reuse rather than a load-bearing proof: the joint-training benefit is demonstrated directly by the TimeLoc comparison in Table 1, and the loss ablations in Table 3 support the TML contribution independently of that citation. Section 21 concedes that Zhai et al., Padilha et al., and Salem et al. lack code or exact splits and that the baselines are reimplementations; that is a reproducibility and fairness risk for the headline comparison, but it is not circularity, because the reimplemented baselines' protocols do not enter the derivation of GT-Loc's own predictions. No equation in the paper equates a predicted output with an input parameter, and no uniqueness claim is imported from the authors' prior work, so no circular step is exhibited.

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

The central results rest on standard Fourier-feature embeddings and on several domain assumptions about how time appears in images. The most important is that temporal visual change is smooth and cyclic, which motivates TML. No new physical entities are introduced. Hand-set scales and noises, plus two learned temperatures, are the main additional parameters beyond the trained MLPs.

free parameters (5)
  • Learnable temperatures for location and time similarity = optimized during training, values not reported
    Used in Eq. 2 and Eq. 7 to scale cosine similarities; trained on data rather than set by theory.
  • Random Fourier feature frequency scales = sigma in {20, 24, 28}
    Chosen by hand following GeoCLIP (Section 8); sets the multi-scale Fourier projection for both location and time inputs.
  • GPS augmentation noise = 150 m in-batch, 1500 m queue
    Hand-chosen augmentation scales for location embeddings, reported in Supplementary Section 8.
  • Time augmentation noise = 0.15 months, 0.15 hours
    Hand-chosen augmentation scales for cyclic time inputs, reported in Supplementary Section 8.
  • Projection and embedding dimensions = image 768 to 512; time/location hidden 1024, output 512
    Architecture capacities chosen by hand; they determine how much temporal and spatial detail can be retained.
assumptions (4)
  • domain assumption Visual appearance varies smoothly with capture time, so similarity targets proportional to temporal difference are a valid learning signal.
    Stated in Section 3.3 as the motivation for TML. If adjacent hours or months are visually discontinuous, the soft targets may not align with the actual feature geometry.
  • domain assumption Frozen CLIP ViT-L/14 features retain enough time-relevant information after MLP projection.
    The image encoder is not fine-tuned (Sections 3 and 8); all temporal prediction capacity is a projection of pretrained CLIP features.
  • domain assumption Night-time and indoor images lack reliable temporal cues and can be removed without biasing the benchmark.
    Supplementary Section 12 filters these images using sunrise/sunset estimates and a Places365 model; no analysis of the filter's effect on the test distribution is provided.
  • standard math Random Fourier features provide a suitable kernel approximation for low-dimensional time and location inputs.
    Used in Eq. 1 and Eq. 5 following Tancik et al.; accepted background from the cited literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space." pith.science (2026). https://pith.science/paper/X2KZU5YY

@misc{pith2026250710473,
  author       = {Pith},
  title        = {Pith review of: GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2KZU5YY}},
  note         = {Machine review of arXiv:2507.10473}
}
read the original abstract

Timestamp prediction aims to determine when an image was captured using only visual information, supporting applications such as metadata correction, retrieval, and digital forensics. In outdoor scenarios, hourly estimates rely on cues like brightness, hue, and shadow positioning, while seasonal changes and weather inform date estimation. However, these visual cues significantly depend on geographic context, closely linking timestamp prediction to geo-localization. To address this interdependence, we introduce GT-Loc, a novel retrieval-based method that jointly predicts the capture time (hour and month) and geo-location (GPS coordinates) of an image. Our approach employs separate encoders for images, time, and location, aligning their embeddings within a shared high-dimensional feature space. Recognizing the cyclical nature of time, instead of conventional contrastive learning with hard positives and negatives, we propose a temporal metric-learning objective providing soft targets by modeling pairwise time differences over a cyclical toroidal surface. We present new benchmarks demonstrating that our joint optimization surpasses previous time prediction methods, even those using the ground-truth geo-location as an input during inference. Additionally, our approach achieves competitive results on standard geo-localization tasks, and the unified embedding space facilitates compositional and text-based image retrieval.

Figures

Figures reproduced from arXiv: 2507.10473 by the authors.

Figure 1
Figure 1. GT-Loc: Our Unified Approach vs. Prior Methods. By mapping image, location and time into a single multimodal em￾bedding space, our method can be used for (a) simultaneous image￾to-location and image-to-time retrieval, (b) composed geotemporal￾to-image retrieval. In contrast, current methods are limited to only (c) location [3, 5], (d) time [27] or (e) geo-temporal classification [45] or (f) image-to-GPS retrieval [1… view at source ↗
Figure 2
Figure 2. Overview of GT-Loc: GT-Loc uses an image encoder V(·), location encoder L (·) and time encoder T (·) to generate a set of image Vi, location Li and time Ti embeddings. Leveraging the CLIP [24] pretrained ViT-L/14 as image encoder, we aim to align its image embedding to both location and time embeddings. The image-location alignment is learned through a regular CLIP-like loss [37] and the image-time alignment is lear… view at source ↗
Figure 3
Figure 3. (a) Proposed Temporal Metric Learning loss Ltime: We compute the image-time similarity matrix by taking the cosine distance between the image and time embedding of the all instances of the batch. We then obtain the target distance matrix by comput￾ing the cyclic toroidal time difference between each pair. As shown in the red highlighted box, we take the ith row of both matrices and normalize them using the softmax f… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: (a) Sample images for two cameras of the SkyFinder [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative examples of geo-localization and time-of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: GT-Loc during inference. Our framework consists of a model that can predict both the location and capture-time of an image at the same time using a retrieval approach. Given a query image I Q, a gallery of GPS coordinates and a gallery of timestamps, GT-Loc maps the th…
Figure 7
Figure 7. Figure 7: Additional qualitative examples of geo-localization and time-of-capture prediction using text queries. Top: prompt passed to [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: (a) Additional sample predictions for three cameras of the SkyFinder test set with the ground truth location and capture time. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: shows two qualitative results of our composi￾tional retrieval model. GPS: (38.4579, -109.8201) Time: Feb 1, 8:39 AM GPS: (38.4579, -109.8201) Time: Feb 7, 7:39 AM GPS: (46.91670, 7.4670) Time: May 4, 7:06 PM GPS: (46.91670, 7.4670) Time: Jul 17, 5:06 PM (a) (b) [PITH_…
Figure 10
Figure 10. Figure 10: Cumulative geolocation error at different thresholds. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 13
Figure 13. Figure 13: t-SNE visualization of image embeddings from non [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 12
Figure 12. Figure 12: PCA plots of the embedding spaces in GT-Loc. (a) [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 15
Figure 15. Figure 15: Sample predictions where GT-Loc outperforms the [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 14
Figure 14. Figure 14: Qualitative time-of-day results near sunrise and sunset. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 65 canonical work pages

  1. [1]

    OpenStreetView-5M: The many roads to global visual ge- olocation

    Guillaume Astruc, Nicolas Dufour, Ioannis Siglidis, Con- stantin Aronssohn, Nacim Bouia, Stephanie Fu, Romain Loiseau, Van Nguyen Nguyen, Charles Raude, Elliot Vincent, Lintao Xu, Hongyu Zhou, and Loic Landrieu. OpenStreetView-5M: The many roads to global visual ge- olocation. CVPR, 2024. 15, 16

    work page 2024

  2. [2]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020. 2, 7

    work page 2020

  3. [3]

    Where we are and what we’re looking at: Query based worldwide image geo-localization using hierarchies and scenes

    Brandon Clark, Alec Kerrigan, Parth Parag Kulkarni, Vi- cente Vivanco Cepeda, and Mubarak Shah. Where we are and what we’re looking at: Query based worldwide image geo-localization using hierarchies and scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23182–23190, 2023. 1, 6

    work page 2023

  4. [4]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180–15190, 2023. 16

    work page 2023

  5. [5]

    Pigeon: Predicting image geolocations

    Lukas Haas, Michal Skreta, Silas Alberti, and Chelsea Finn. Pigeon: Predicting image geolocations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12893–12902, 2024. 1, 3, 6

    work page 2024

  6. [6]

    Deep outdoor illumination estimation

    Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-Fran c ¸ois Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 3, 18

    work page 2017

  7. [7]

    Openclip, 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Ha- jishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 7, 13

    work page 2021

  8. [8]

    Geolocating static cameras

    Nathan Jacobs, Scott Satkin, Nathaniel Roman, Richard Speyer, and Robert Pless. Geolocating static cameras. In 2007 IEEE 11th International Conference on Computer Vi- sion, pages 1–6. IEEE, 2007. 3, 17

    work page 2007

Show all 68 references
  1. [9]

    Braswell, Andrew D

    Nathan Jacobs, Walker Burgin, Nick Fridrich, Austin Abrams, Kylia Miskell, Bobby H. Braswell, Andrew D. Richardson, and Robert Pless. The global network of outdoor webcams: Properties and applications. In ACM SIGSPATIAL Interna- tional Conference on Advances in Geographic Info...

    2009

  2. [10]

    Time2vec: Learning a vector representation of time

    Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus Brubaker. Time2vec: Learning a vector representation of time. arXiv preprint arXiv:1907.05321, 2019. 7, 13

    1907 arXiv

  3. [11]

    Satclip: Global, general- purpose location embeddings with satellite imagery

    Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, general- purpose location embeddings with satellite imagery. arXiv preprint arXiv:2311.17179, 2023. 1, 2, 3, 7, 16

    2023 arXiv

  4. [12]

    Cityguessr: City-level video geo-localization on a global scale

    Parth Parag Kulkarni, Gaurav Kumar Nayak, and Mubarak Shah. Cityguessr: City-level video geo-localization on a global scale. In European Conference on Computer Vision, pages 293–311. Springer, 2024. 3

    2024

  5. [13]

    What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24–51, 2010

    Jean-Franc ¸ois Lalonde, Srinivasa G Narasimhan, and Alexei A Efros. What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24–51, 2010. 3

    2010

  6. [14]

    Estimating the natural illumination conditions from a single outdoor image

    Jean-Franc ¸ois Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image. International Journal of Com- puter Vision, 98:123–145, 2012. 3, 18

    2012

  7. [15]

    The benchmarking initiative for multimedia evaluation: Mediaeval 2016

    Martha Larson, Mohammad Soleymani, Guillaume Gravier, Bogdan Ionescu, and Gareth JF Jones. The benchmarking initiative for multimedia evaluation: Mediaeval 2016. IEEE MultiMedia, 24(1):93–96, 2017. 6

    2016

  8. [16]

    Are you lying: Validating the time-location of outdoor images

    Xiaopeng Li, Wenyuan Xu, Song Wang, and Xianshan Qu. Are you lying: Validating the time-location of outdoor images. In Applied Cryptography and Network Security: 15th Interna- tional Conference, ACNS 2017, Kanazawa, Japan, July 10-12, 2017, Proceedings 15, pages 103–123. Sprin...

    2017

  9. [17]

    Presence- only geographical priors for fine-grained image classification

    Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence- only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3, 7, 13

    2019

  10. [18]

    CSP: Self-supervised contrastive spatial pre- training for geospatial-visual representations

    Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Ste- fano Ermon. CSP: Self-supervised contrastive spatial pre- training for geospatial-visual representations. In Proceedings of the 40th International Conference on Machine Learning, pages 23498–23515. PMLR, 2023. 2, 3, 16

    2023

  11. [19]

    Sky segmentation in the wild: An empir- ical study

    Radu Paul Mihail, Scott Workman, Zach Bessinger, and Nathan Jacobs. Sky segmentation in the wild: An empir- ical study. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1–6, 2016. 6

    2016

  12. [20]

    Geolocation estimation of photos using a hierarchical model and scene classification

    Eric Muller-Budack, Kader Pustu-Iren, and Ralph Ewerth. Geolocation estimation of photos using a hierarchical model and scene classification. In Proceedings of the European conference on computer vision (ECCV), pages 563–579, 2018. 3, 6

    2018

  13. [21]

    Maxime Oquab, Timoth´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicola...

  14. [22]

    Content-aware detection of temporal metadata manipulation

    Rafael Padilha, Tawfiq Salem, Scott Workman, Fernanda A Andal´o, Anderson Rocha, and Nathan Jacobs. Content-aware detection of temporal metadata manipulation. IEEE Transac- tions on Information Forensics and Security, 17:1316–1327,

  15. [23]

    Where in the world is this image? transformer-based geo-localization in the wild

    Shraman Pramanick, Ewa M Nowara, Joshua Gleason, Car- los D Castillo, and Rama Chellappa. Where in the world is this image? transformer-based geo-localization in the wild. In European Conference on Computer Vision, pages 196–215. Springer, 2022. 3, 6

    2022

  16. [24]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  17. [25]

    Bridging the domain gap for ground-to-aerial image matching

    Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 470–479, 2019. 3

    2019

  18. [26]

    Learning a dynamic map of visual appearance

    Tawfiq Salem, Scott Workman, and Nathan Jacobs. Learning a dynamic map of visual appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435–12444, 2020. 2, 6, 17, 18

    2020

  19. [27]

    Timestamp estimation from outdoor scenes

    Tawfiq Salem, Jisoo Hwang, and Rafael Padilha. Timestamp estimation from outdoor scenes. In Annual ADFSL Con- ference on Digital Forensics, Security and Law , number 2. Embry-Riddle Aeronautical University Commons, 2022. 1, 2, 6, 18

    2022

  20. [28]

    Cplanet: Enhancing image geolocalization by combina- torial partitioning of maps

    Paul Hongsuck Seo, Tobias Weyand, Jack Sim, and Bohyung Han. Cplanet: Enhancing image geolocalization by combina- torial partitioning of maps. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 536–551,

  21. [29]

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

    Yujiao Shi, Liu Liu, Xin Yu, and Hongdong Li. Spatial- aware feature aggregation for image based cross-view geo- localization. Advances in Neural Information Processing Systems, 32, 2019. 3

    2019

  22. [30]

    Where am i looking at? joint location and orientation estima- tion by cross-view matching

    Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estima- tion by cross-view matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4064–4072, 2020. 3

    2020

  23. [31]

    Everything at once-multi- modal fusion transformer for video retrieval

    Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multi- modal fusion transformer for video retrieval. In Proceedings of the ieee/cvf conference on computer vision ...

    2022

  24. [32]

    Preserving modality struc- ture improves multi-modal learning

    Sirnam Swetha, Mamshad Nayeem Rizve, Nina Shvetsova, Hilde Kuehne, and Mubarak Shah. Preserving modality struc- ture improves multi-modal learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 21993–22003, 2023. 8, 16

    2023

  25. [33]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information pr...

    2020

  26. [34]

    Yfcc100m: The new data in multimedia research

    Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li- Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64–73, 2016. 6, 17

    2016

  27. [35]

    Coming down to earth: Satellite-to-street view syn- thesis for geo-localization

    Aysim Toker, Qunjie Zhou, Maxim Maximov, and Laura Leal- Taix´e. Coming down to earth: Satellite-to-street view syn- thesis for geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6488–6497, 2021. 3

    2021

  28. [36]

    Photo sundial: estimating the time of capture in consumer photos

    Tsung-Hung Tsai, Wei-Cih Jhou, Wen-Huang Cheng, Min- Chun Hu, I-Chao Shen, Tekoing Lim, Kai-Lung Hua, Ahmed Ghoneim, M Anwar Hossain, and Shintami C Hidayati. Photo sundial: estimating the time of capture in consumer photos. Neurocomputing, 177:529–542, 2016. 2, 18

    2016

  29. [37]

    Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization

    Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. Ad- vances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6, 7, 16

    2024

  30. [38]

    Revisiting im2gps in the deep learning era

    Nam V o, Nathan Jacobs, and James Hays. Revisiting im2gps in the deep learning era. In Proceedings of the IEEE inter- national conference on computer vision, pages 2621–2630,

  31. [39]

    Center for international earth science information net- work—ciesin—columbia university.(2016)

    L Warszawski, K Frieler, V Huber, F Piontek, O Serdeczny, X Zhang, Q Tang, M Pan, Y Tang, Q Tang, et al. Center for international earth science information net- work—ciesin—columbia university.(2016). gridded popula- tion of the world, version 4 (gpwv4): Population density. pa...

    2016

  32. [40]

    Shadow detection and sun direction in photo collections

    Scott Wehrwein, Kavita Bala, and Noah Snavely. Shadow detection and sun direction in photo collections. In 2015 International Conference on 3D Vision, pages 460–468. IEEE,

    2015

  33. [41]

    Planet- photo geolocation with convolutional neural networks

    Tobias Weyand, Ilya Kostrikov, and James Philbin. Planet- photo geolocation with convolutional neural networks. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceed- ings, Part VIII 14, pages 37–55. Springer, 2016. 3, 6

    2016

  34. [42]

    Wide- area image geolocalization with aerial reference imagery

    Scott Workman, Richard Souvenir, and Nathan Jacobs. Wide- area image geolocalization with aerial reference imagery. In IEEE International Conference on Computer Vision (ICCV), pages 1–9, 2015. Acceptance rate: 30.3%. 17

    2015

  35. [43]

    Mind the modality gap: Towards a remote sensing vision-language model via cross-modal alignment

    Angelos Zavras, Dimitrios Michail, Beg¨um Demir, and Ioan- nis Papoutsis. Mind the modality gap: Towards a remote sensing vision-language model via cross-modal alignment. arXiv preprint arXiv:2402.09816, 2024. 3

    2024 arXiv

  36. [44]

    Rank-n-contrast: Learning continuous representations for regression

    Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: Learning continuous representations for regression. Advances in Neural Information Processing Systems, 36, 2024. 7 10

    2024

  37. [45]

    Learning geo-temporal image features

    Menghua Zhai, Tawfiq Salem, Connor Greenwell, Scott Work- man, Robert Pless, and Nathan Jacobs. Learning geo-temporal image features. In British Machine Vision Conference, 2019. 1, 2, 6, 8, 15, 17

    2019

  38. [46]

    Places: A 10 million image database for scene recognition

    Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 12

    2017

  39. [47]

    Img2loc: Re- visiting image geolocalization using multi-modality founda- tion models and image-based retrieval-augmented generation

    Zhongliang Zhou, Jielu Zhang, Zihan Guan, Mengxuan Hu, Ni Lao, Lan Mu, Sheng Li, and Gengchen Mai. Img2loc: Re- visiting image geolocalization using multi-modality founda- tion models and image-based retrieval-augmented generation. In Proceedings of the 47th International ACM ...

    2024

  40. [48]

    Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment

    Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 16

    arXiv

  41. [49]

    Vigor: Cross- view image geo-localization beyond one-to-one retrieval

    Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross- view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3640–3649, 2021. 3

    2021

  42. [50]

    Transgeo: Trans- former is all you need for cross-view image geo-localization

    Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Trans- former is all you need for cross-view image geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162–1171, 2022. 3 11 GT-Loc: Unifying When and Where in Images Th...

    2022

  43. [51]

    GT-Loc during inference

    Model inference Figure 6. GT-Loc during inference. Our framework consists of a model that can predict both the location and capture-time of an image at the same time using a retrieval approach. Given a query imageI Q, a gallery of GPS coordinates and a gallery of timestamps, G...

  44. [52]

    We use the same architecture for the time and location encoders as GeoCLIP

    Implementation details Following GeoCLIP, the backbone of the image encoder is a pretrained ViT-L/14 from CLIP and the MLP consists of two fully connected layers with the ReLU activation function and dimensions 768 and 512 respectively. We use the same architecture for the tim...

  45. [53]

    We use Adam optimizer with coefficients β1 = 0 .9, β2 = 0 .999 and ℓ2 penalty of 1 × 10−6

    Training protocol GT-Loc is trained for 20 epochs using a cosine decay scheduler, with learning rate values ranging from αmax = 3 × 10−5 to αmin = 3 × 10−7. We use Adam optimizer with coefficients β1 = 0 .9, β2 = 0 .999 and ℓ2 penalty of 1 × 10−6. For the contrastive losses, w...

  46. [54]

    We include a failure case, on the last row of figure 8, where the time error is high because of the presence of fog in the image

    Additional qualitative results We show additional qualitative results of our method in fig- ures 7 and 8. We include a failure case, on the last row of figure 8, where the time error is high because of the presence of fog in the image

  47. [55]

    Time-of-capture prediction histograms The time prediction histograms, shown in figures 4, 5, 7, and 8, are computed using the following equation: Ci = NGX j=1 1 [j∈Bi] · I Q · T G j , (12) where Bi is the set of gallery embeddings that correspond to the ith bin, Ci is the bin ...

  48. [56]

    In partic- ular, we remove all night-time and indoor images, since they often have inconsistent temporal cues

    Dataset details We apply two filters to remove samples from the CVT that don’t provide meaningful temporal information. In partic- ular, we remove all night-time and indoor images, since they often have inconsistent temporal cues. To remove night images, we estimate the sunris...

  49. [57]

    Additional ablations 13.1. Image backbones To evaluate the impact of different image embeddings on time prediction performance, we conducted ablation studies using three backbones: DINOv2-L [ 21], OpenCLIP ViT- G [7], and OpenAI’s original CLIP ViT-L [24]. For these experiment...

    1901

  50. [58]

    [45] for compositional retrieval

    Compositional image retrieval details To compare GT-Loc against a suitable baseline, we repurpose the model proposed by Zhai et al. [45] for compositional retrieval. In its original form, the model comprises an im- age encoder CI (I), a time encoder CT (t), and a location enco...

  51. [59]

    In addition to our main results, we include evaluations on two extra datasets: YFCC26k and OSV-5M

    Additional geo-localization analysis Figure 10 presents the cumulative geolocation error evalu- ated over a range of distance thresholds. In addition to our main results, we include evaluations on two extra datasets: YFCC26k and OSV-5M. These plots allow for a more com- prehen...

  52. [60]

    Effect of population density on geo- localization Geo-localization datasets, particularly those collected via web-scraping or social media platforms, are inherently bi- ased toward regions with higher population density. This raises a natural question: how does the performance...

  53. [61]

    gallery size Figure 11 shows how the size of the gallery affects geo- localization and time prediction performance

    Model performance vs. gallery size Figure 11 shows how the size of the gallery affects geo- localization and time prediction performance. We observe that increasing the size of the gallery leads to improved re- sults, but the gains tend to saturate quickly. In particular, a ga...

  54. [62]

    Analysis of the learned embedding space One of the key motivations for GT-Loc is to align images, time, and location in a shared multimodal embedding space. This approach is inspired by prior works like GeoCLIP [37], SatCLIP [11], and CSP [18], which embed images and GPS coord...

  55. [63]

    We begin by randomly selecting 20 panoramas from the CVUSA dataset [ 42]

    Embedding distribution of non-overlapping panorama crops To further analyze the structure of our learned embedding space, we conduct a qualitative visualization using t-SNE (Figure 13). We begin by randomly selecting 20 panoramas from the CVUSA dataset [ 42]. From each panoram...

  56. [64]

    t-SNE visualization of image embeddings from non- overlapping 90◦ crops of 20 panoramas sampled from the CVUSA dataset

    Scalability of the retrieval approach In Table 11, we present a comparison of the memory usage and computational cost (measured in FLOPs) between clas- sification and regression baselines and our retrieval-based Figure 13. t-SNE visualization of image embeddings from non- over...

  57. [65]

    For instance, Zhai et al

    Limitations and reproducibility challenges of existing time prediction methods Most previous time prediction methods suffer from a lack of standardization in their training and evaluation protocols. For instance, Zhai et al. [45] use subsets of the AMOS [8, 9] and YFCC100M [34...

  58. [66]

    Their use of top-k accuracy as an evaluation metric further complicates direct comparisons

    also omit critical details such as dataset splits for the SkyFinder dataset, do not clarify whether their results cor- responds to same- or cross-camera evaluation, and fail to provide the source code for replication. Their use of top-k accuracy as an evaluation metric further...

  59. [67]

    Qualitative time-of-day results near sunrise and sunset Predicting the time of day during periods close to sunrise and sunset is particularly challenging due to the visual similarity of scenes captured around these times. The task is further complicated by its strong dependenc...

  60. [68]

    The results suggest that TimeLoc struggles more with hour predictions at higher latitudes (40° to 70°) compared to GT-Loc

    GT-Loc predictions across different lati- tudes Figure 15 compares time prediction examples from GT-Loc and TimeLoc, a baseline model trained solely with the vi- sual (V) and temporal ( T ) encoders. The results suggest that TimeLoc struggles more with hour predictions at high...

Pith tools

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