Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Towards Interpretable Geo-localization: a Concept-Aware Global Image-GPS Alignment Framework

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A concept bottleneck between images and GPS coordinates makes geo-localization explainable without sacrificing accuracy.

desk verdict A plausible concept-bottleneck extension of GeoCLIP with small, consistent accuracy gains, but the interpretability story needs grounding before it can carry the paper. read the letter →

arxiv 2509.01910 v2 pith:GGDOLSIL submitted 2025-09-02 cs.CV cs.AI

classification cs.CVcs.AI
keywords geo-localizationinterpretabilityconceptbottleneckCLIPimage-GPSalignmentgeographicconceptscontrastivelearninglocationembeddings
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

This paper proposes making image geo-localization — predicting where on Earth a photo was taken — answerable in human terms by inserting a concept bottleneck between images, GPS coordinates, and the final location prediction. The authors claim this is the first framework to bring concept-based interpretability to geo-localization, and they show the added concept layer does not cost accuracy: on the Im2GPS3k benchmark it improves over the GeoCLIP baseline at every distance threshold, with the largest gain at street level (13.2% vs 10.8% within 1 km). The mechanism is a shared "concept subspace" built on a manually curated set of geographic concepts, in which image and location embeddings are aligned, and it lets the model report which concepts (e.g., skyscraper, typhoon, historical architecture) drive a prediction. The paper also reports gains on downstream geospatial tasks and evidence that the learned location embeddings line up with real geographic distributions, such as U.S. forest cover.

What carries the argument

The load-bearing object is the Concept-Aware Alignment Module with a learnable concept basis B ∈ R^{d×k}, each column starting from the frozen CLIP text embedding of a named geographic concept (B = E_concept + Δ). Image features are mapped into the concept subspace through a lightweight MLP and location features are projected directly onto B, so both modalities are compared in a space whose axes have semantic names. A contrastive image-to-location loss aligns matched pairs, while a Gaussian-kernel divergence loss enforces distribution-level consistency between image and location concept scores. This dual objective is what the paper claims simultaneously transfers world knowledge into the loc

What would settle it

Measure the cosine similarity between each trained column of B and its source CLIP text embedding; if a column's nearest neighbor in the concept vocabulary is no longer its own concept name, the semantic grounding claim fails. Also, retrain with a shuffled concept set: if accuracy and purported explanations do not degrade, the Geography-Driven Concept Set is not carrying the effect.

Watch

Extended reading notes

Core claim

The central discovery is that a CLIP-based image-GPS contrastive model can be trained with a concept bottleneck as an intermediate layer without sacrificing retrieval accuracy; in fact the concept-aware alignment improves it. The method constructs a Geography-Driven Concept Set from geographic knowledge sources, enriches it with an LLM, and manually filters it, then encodes concepts with a frozen CLIP text encoder to form a basis matrix B = E_concept + Δ. Image features are projected into the resulting concept subspace with an MLP, location features by x_loc^T B, and training minimizes an image-to-location contrastive loss plus a distribution-level divergence loss that pulls the two modaliti

Load-bearing premise

Interpretability rests on the assumption that after training, each axis of the concept subspace still means the geographic concept named by its text embedding — the learnable offset Δ in Eq. 1 is never constrained or measured, and the curated concept set is not released.

Editorial extensions

If this is right

  • If the claim holds, geo-localization models can output a ranked list of geographic concepts explaining each prediction, not just a point estimate.
  • Location embeddings become queryable for concept similarity (e.g., "forest"), producing maps that correlate with real-world distributions.
  • Concept-aware pretraining transfers to tasks beyond retrieval: income, education, temperature, species classification, and country classification improve over GeoCLIP embeddings.
  • The improvement is largest at fine scales (1 km, 25 km), suggesting concept cues matter most when visual scenes look similar across regions.
  • The concept subspace itself becomes the interface for explanations, enabling individual and global interpretability from the same representation.

Reading between the lines

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

  • The paper does not constrain the learnable offset Δ in Eq. 1, so one testable extension is to measure how far the trained concept basis drifts from its source text embeddings; if columns change meaning, explanations could inherit the text labels while encoding something else.
  • Because the Geography-Driven Concept Set is manually curated and not released, the framework's reproducibility depends on the concept bank; an automated, released concept-selection procedure would let others test whether concept composition rather than the alignment loss drives the gains.
  • The concept-level explanations shown are correlational (activation medians, post-hoc CBM weights), not causal; a natural extension is to intervene on concept scores and verify that the predicted location shifts in the direction the concept names imply.
  • The same concept-bottleneck alignment recipe could be applied to other spatially grounded vision-language tasks, such as satellite image retrieval or field-level biodiversity mapping, where interpretable geographic cues are also valuable.
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 / 4 minor

Summary. The paper proposes an interpretable geo-localization framework that augments GeoCLIP-style image–GPS contrastive learning with a concept bottleneck. A manually/LLM-constructed Geography-Driven Concept Set is encoded with a frozen CLIP text encoder to form initial concept directions; a learnable offset Δ is added to form the concept basis B (Eq. 1). Image features are projected via an MLP into a k-dimensional concept space (Eq. 2), location features via B (Eq. 3), and training combines an image-to-GPS contrastive loss with a kernel-based concept-space divergence loss (Eqs. 4–7). Evaluations on Im2GPS3k report accuracy improvements over GeoCLIP (e.g., +2.4 points at 1 km), gains on downstream geospatial tasks, and qualitative/quantitative interpretability analyses (concept contributions, global decision rules, concept activation rankings, and concept similarity maps). The paper claims to be the first to introduce interpretability to geo-localization. The central accuracy claim is benchmarked externally, but the interpretability claims rest on the semantic identity of the learned concept directions, which is not directly verified.

Significance. If the claims are validated, the concept-aware alignment idea is a meaningful contribution: it connects an established contrastive geo-localization objective to concept bottlenecks, reports consistent accuracy gains on an external benchmark, and quantifies concept–geography alignment (e.g., forest similarity correlation ρ = 0.6525 vs. GeoCLIP's 0.3536). The paper's strengths include clear architecture specification, ablations against a general concept set (SpLiCE), downstream-task evaluation, and explicit threshold-based metrics. However, the core interpretability claim depends on the learned concept basis remaining semantically identified with the named concepts after training; the paper does not yet provide evidence for this. The 'first work' novelty claim is also contradicted by the paper's own citation of Theiner et al. 2022. These issues are fixable within the scope of the manuscript, but they currently prevent acceptance.

major comments (4)
  1. [Concept-Aware Alignment Module, Eq. (1)] The central interpretability claim is not yet grounded. B = E_concept + Δ, with Δ unconstrained and never measured. The loss (Eq. 7) does not prevent Δ from rotating/rescaling the concept directions so that columns of B no longer correspond to the named concepts. Since z_img is produced by an arbitrary MLP f_img (Eq. 2) and z_loc = x_loc^T B, the dimensions interpreted in Figs. 3–5 and Table 3 may not be the concepts they are labeled as. Please report ||Δ||_F, per-concept cosine similarity between columns of B and E_concept, or a nearest-neighbor check after training; alternatively, add a regularization or projection constraint to enforce semantic stability.
  2. [Abstract and Introduction] The statement 'this is the first work to introduce interpretability into geo-localization' is contradicted by the paper's own reference to Theiner et al. 2022, 'Interpretable Semantic Photo Geolocation,' which is an interpretability-oriented geo-localization method using semantic partitioning. The claim should be qualified, e.g., 'first concept-bottleneck interpretability for image–GPS alignment' or similar, to avoid a clear novelty overclaim.
  3. [Table 4 and Table 5] Table 4 reports SpLiCE accuracy at 200 km as 33.16, which is barely above its 25 km value (32.92) and far below the Geography-Driven value at 200 km (49.8); the 750 km value (66.73) is also anomalously low relative to the pattern. This suggests a likely typo or evaluation inconsistency that undermines the concept-set ablation. Table 5's 'W/o Concept-Aware Alignment Module' row exactly matches the GeoCLIP row in Table 1 at all thresholds, yet no error bars, number of runs, or significance tests are provided anywhere. Please re-check the SpLiCE numbers and report variances across multiple seeds.
  4. [Concept-Aware Embedding Analysis, Fig. 5] The quantitative forest-correlation analysis uses the same concept text embeddings (E_concept) that the model was trained to align with, and the correlation is computed on state-level averages. Because B = E_concept + Δ is trained, the reported correlation may partly reflect training to the concept set rather than an emergent geographic semantic. Please validate with held-out concepts (concepts not used in training) or a model trained with a different concept set, and report the correlation using the trained B columns as well as E_concept. Additionally, the Geography-Driven Concept Set is not released, so independent verification of concept coverage is impossible; please include the full set and the LLM prompt as supplementary material.
minor comments (4)
  1. [Method, Section 3] Typo: 'We integrates a Geography-Driven Concept Set' should be 'We integrate'. Also, 'Effectness' in both ablation subsection headings should be 'Effectiveness'.
  2. [Experiments, Dataset and Implementation Details] Capitalize consistently: 'Ten Crop' should be 'ten-crop' and 'IM2GPS3k' appears with inconsistent capitalization (Im2GPS3k elsewhere). Minor grammar: 'we also uses a Ten Crop strategy' → 'we also use a ten-crop strategy'.
  3. [Table 3 caption] The caption says 'median scores (ms)', which can be confused with milliseconds; consider 'med. score' and 'lowest-8' → 'bottom-8' for readability.
  4. [Supplementary material] The LLM prompt used for concept enrichment and the full Geography-Driven Concept Set are mentioned as included in the supplementary material; they are not present in the reviewed version. Please include them, as they are essential for reproducibility of the concept construction pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy claims are evaluated on an external benchmark (Im2GPS3k) and the main interpretability evidence includes an external ground-truth comparison.

full rationale

The paper's central geo-localization accuracy claim is benchmarked externally on Im2GPS3k against GeoCLIP and other published methods; no fitted parameter from the test set is renamed as a prediction. The interpretability claims are supported partly by the model's own concept-space activations, but the key quantitative interpretability result (Figure 5) correlates model location embeddings with external U.S. forest coverage data (ρ = 0.6525, p < 0.001), so it is not circular by construction. The concept-importance rankings (Table 3, Figure 4) are descriptive summaries of the model's own behavior, not predictions forced by the loss. Two non-circular weaknesses exist, but they are not circularity: (1) the abstract's 'first work to introduce interpretability into geo-localization' is contradicted by the paper's own citation of Theiner et al. 2022, 'Interpretable Semantic Photo Geolocation'; (2) Eq. 1 defines B = E_concept + Δ with no reported bound or measurement of Δ, so the semantic stability of the trained concept directions is assumed rather than demonstrated. These are correctness/validity concerns, not cases where the derivation reduces to its inputs. The external benchmark and external ground-truth correlation keep the central claims self-contained, so a circularity score of 0 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim depends on a hand-built concept set, an unconstrained learnable basis, and several unspecified hyperparameters; the paper ships no code or concept list.

free parameters (5)
  • lambda = 10
    Weight for concept divergence loss in Eq. 7, chosen without reported sensitivity analysis.
  • tau = learnable
    Temperature in contrastive loss Eq. 4.
  • Delta = learned offset matrix
    Offset added to CLIP concept embeddings in Eq. 1; unconstrained, so concept basis may drift from semantic meanings.
  • k = not specified
    Number of concepts in Geography-Driven Concept Set; not stated in the text.
  • sigma = not specified
    Bandwidth of the Gaussian kernel in Eq. 5; not stated or analyzed.
assumptions (4)
  • domain assumption CLIP text encodings of geographic concepts carry the semantics needed for interpretable geo-alignment
    The whole architecture projects onto CLIP text embeddings (Eqs. 1-3).
  • domain assumption GeoCLIP location encoder provides an adequate base embedding space
    The location encoder is taken from GeoCLIP and only its output is projected.
  • ad hoc to paper The manually and LLM constructed concept set covers the discriminative cues for worldwide geo-localization
    The concept set is curated specifically for this paper and not released, so its coverage cannot be independently checked.
  • ad hoc to paper The concept subspace remains interpretable after learning Delta
    No constraint or drift measurement on B in Eq. 1.
invented entities (1)
  • Geography-Driven Concept Set
    purpose: A fixed bank of geographic concepts used as an interpretable intermediate subspace for image-GPS alignment.
    The set is not released; the prompt used for LLM generation is only in the supplementary material. No external handle to falsify the set itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Interpretable Geo-localization: a Concept-Aware Global Image-GPS Alignment Framework." pith.science (2026). https://pith.science/paper/GGDOLSIL

@misc{pith2026250901910,
  author       = {Pith},
  title        = {Pith review of: Towards Interpretable Geo-localization: a Concept-Aware Global Image-GPS Alignment Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGDOLSIL}},
  note         = {Machine review of arXiv:2509.01910}
}
read the original abstract

Worldwide geo-localization involves determining the exact geographic location of images captured globally, typically guided by geographic cues such as climate, landmarks, and architectural styles. Despite advancements in geo-localization models like GeoCLIP, which leverages images and location alignment via contrastive learning for accurate predictions, the interpretability of these models remains insufficiently explored. Current concept-based interpretability methods fail to align effectively with Geo-alignment image-location embedding objectives, resulting in suboptimal interpretability and performance. To address this gap, we propose a novel framework integrating global geo-localization with concept bottlenecks. Our method inserts a Concept-Aware Alignment Module that jointly projects image and location embeddings onto a shared bank of geographic concepts (e.g., tropical climate, mountain, cathedral) and minimizes a concept-level loss, enhancing alignment in a concept-specific subspace and enabling robust interpretability. To our knowledge, this is the first work to introduce interpretability into geo-localization. Extensive experiments demonstrate that our approach surpasses GeoCLIP in geo-localization accuracy and boosts performance across diverse geospatial prediction tasks, revealing richer semantic insights into geographic decision-making processes.

Figures

Figures reproduced from arXiv: 2509.01910 by the authors.

Figure 1
Figure 1. Concept-aware geo-localization overview. (A) CLIP-based geo-localization model predicts GPS coordi￾nates by directly mapping visual embeddings to a location embedding gallery, following the image-location alignment. (B) Our approach inserts a concept bottleneck between im￾age and GPS, grounding GPS prediction in semantically meaningful cues. retrieval, it still faces the challenge of distinguishing visu￾ally similar… view at source ↗
Figure 2
Figure 2. The pretraining and inference pipeline. Left: The pretraining framework with a Concept-Aware Alignment Module integrated into the image–GPS contrastive learning process. Right: The inference process of image geo-localization. The overall architecture is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of concept contributions in country-level geo-localization task with Label-free CBM [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Global decision rules in image geo-localization. Diagram showing prominent visual concepts differentiat￾ing image geo-localization predictions between China and Japan. Examples highlight region-specific concepts illustrat￾ing how the model’s learned concepts reflect me…
Figure 5
Figure 5. Figure 5: D, our model’s similarity map achieves a higher cor￾relation (ρ = 0.6525, p < 0.001) compared to GeoCLIP (ρ = 0.3536, p = 0.013), indicating enhanced representa￾tion of geographic concepts in our location embeddings. To further answer RQ4 and explore whether concept￾aw…
Figure 6
Figure 6. Figure 6: UMAP visualization of image embeddings from the Im2GPS3k dataset learned by the Concept-Aware Align￾ment Module. Each point represents an image colored by k￾means cluster assignment. More visualizations are provided in the supplementary materials. localization performa…
Figure 7
Figure 7. Figure 7: UMAP visualization of embeddings from im￾ages, locations, and concepts. We project image–location embedding pairs and concept text embeddings (from CLIP) into 2D using UMAP, with lines connecting each image to its corresponding location. Compared to GeoCLIP, our model …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. CarbonCLIP: Enhance Carbon Prediction from Satellite Imagery via Integrated Street-View Semantics and Temporal Context Training

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A dual-branch contrastive learning framework distills street-view semantics and temporal context into satellite representations, improving monthly carbon emission prediction using only satellite imagery at inference.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

  2. [2]

    Beery, S.; Wu, G.; Edwards, T.; Pavetic, F.; Majewski, B.; Mukherjee, S.; Chan, S.; Morgan, J.; Rathod, V.; and Huang, J. 2022. The auto arborist dataset: a large-scale benchmark for multiview urban forest monitoring under domain shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21294--21307

  3. [3]

    Bhalla, U.; Oesterling, A.; Srinivas, S.; Calmon, F.; and Lakkaraju, H. 2024. Interpreting clip with sparse linear concept embeddings (splice). Advances in Neural Information Processing Systems, 37: 84298--84328

  4. [4]

    P.; Cepeda, V

    Clark, B.; Kerrigan, A.; Kulkarni, P. P.; Cepeda, V. V.; and Shah, M. 2023. 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, 23182--23190

  5. [5]

    Dsouza, A.; Tempelmeier, N.; Yu, R.; Gottschalk, S.; and Demidova, E. 2021. Worldkg: A world-scale geographic knowledge graph. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 4475--4484

  6. [6]

    Y.; and Kim, B

    Ghorbani, A.; Wexler, J.; Zou, J. Y.; and Kim, B. 2019. Towards automatic concept-based explanations. Advances in Neural Information Processing Systems, 32

  7. [7]

    Hays, J.; and Efros, A. A. 2008. Im2gps: estimating geographic information from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1--8. IEEE

  8. [8]

    Hooker, J.; Duveiller, G.; and Cescatti, A. 2018. A global dataset of air temperature derived from satellite remote sensing and weather stations. Scientific data, 5(1): 1--11

Show all 48 references
  1. [9]

    Jia, P.; Liu, Y.; Li, X.; Zhao, X.; Wang, Y.; Du, Y.; Han, X.; Wei, X.; Wang, S.; and Yin, D. 2024. G3: an effective and adaptive framework for worldwide geolocalization using large multi-modality models. Advances in Neural Information Processing Systems, 37: 53198--53221

  2. [10]

    Jia, P.; Park, S.; Gao, S.; Zhao, X.; and Li, Y. 2025. GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization. arXiv preprint arXiv:2505.13731

  3. [11]

    Klemmer, K.; Rolf, E.; Robinson, C.; Mackey, L.; and Ru wurm, M. 2025. Satclip: Global, general-purpose location embeddings with satellite imagery. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 4347--4355

  4. [12]

    W.; Nguyen, T.; Tang, Y

    Koh, P. W.; Nguyen, T.; Tang, Y. S.; Mussmann, S.; Pierson, E.; Kim, B.; and Liang, P. 2020. Concept bottleneck models. In International conference on machine learning, 5338--5348. PMLR

  5. [13]

    Larson, M.; Soleymani, M.; Gravier, G.; Ionescu, B.; and Jones, G. J. 2017. The benchmarking initiative for multimedia evaluation: MediaEval 2016. IEEE MultiMedia, 24(1): 93--96

  6. [14]

    W.; Zhang, Y.; Kwon, Y.; Yeung, S.; and Zou, J

    Liang, V. W.; Zhang, Y.; Kwon, Y.; Yeung, S.; and Zou, J. Y. 2022. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35: 17612--17625

  7. [15]

    Liu, L.; and Li, H. 2019. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5624--5633

  8. [16]

    Muller-Budack, E.; Pustu-Iren, K.; and Ewerth, R. 2018. Geolocation estimation of photos using a hierarchical model and scene classification. In Proceedings of the European conference on computer vision (ECCV), 563--579

  9. [17]

    M.; and Weng, T.-W

    Oikarinen, T.; Das, S.; Nguyen, L. M.; and Weng, T.-W. 2023. Label-free Concept Bottleneck Models. In The Eleventh International Conference on Learning Representations

  10. [18]

    Parekh, J.; Khayatan, P.; Shukor, M.; Newson, A.; and Cord, M. 2024. A concept-based explainability framework for large multimodal models. Advances in Neural Information Processing Systems, 37: 135783--135818

  11. [19]

    M.; Gleason, J.; Castillo, C

    Pramanick, S.; Nowara, E. M.; Gleason, J.; Castillo, C. D.; and Chellappa, R. 2022. Where in the world is this image? transformer-based geo-localization in the wild. In European Conference on Computer Vision, 196--215. Springer

  12. [20]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

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

  13. [21]

    Roberts, J.; L \"u ddecke, T.; Das, S.; Han, K.; and Albanie, S. 2023. GPT4GEO: How a language model sees the world's geography. arXiv preprint arXiv:2306.00020

  14. [22]

    Ruiz Luyten, M.; and van der Schaar, M. 2024. A theoretical design of concept sets: improving the predictability of concept bottleneck models. Advances in Neural Information Processing Systems, 37: 100160--100195

  15. [23]

    Ru wurm, M.; Wang, S.; Korner, M.; and Lobell, D. 2020. Meta-learning for few-shot land cover classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, 200--201

  16. [24]

    Sathianarayanan, M.; Hsu, P.-H.; and Chang, C.-C. 2024. Extracting disaster location identification from social media images using deep learning. International Journal of Disaster Risk Reduction, 104: 104352

  17. [25]

    H.; Weyand, T.; Sim, J.; and Han, B

    Seo, P. H.; Weyand, T.; Sim, J.; and Han, B. 2018. Cplanet: Enhancing image geolocalization by combinatorial partitioning of maps. In Proceedings of the European Conference on Computer Vision (ECCV), 536--551

  18. [26]

    Shi, Y.; Yu, X.; Campbell, D.; and Li, H. 2020. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4064--4072

  19. [27]

    Theiner, J.; M \"u ller-Budack, E.; and Ewerth, R. 2022. Interpretable semantic photo geolocation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 750--760

  20. [28]

    A.; Friedland, G.; Elizalde, B.; Ni, K.; Poland, D.; Borth, D.; and Li, L.-J

    Thomee, B.; Shamma, D. A.; Friedland, G.; Elizalde, B.; Ni, K.; Poland, D.; Borth, D.; and Li, L.-J. 2016. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2): 64--73

  21. [29]

    Tian, Y.; Chen, C.; and Shah, M. 2017. Cross-view image matching for geo-localization in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3608--3616

  22. [30]

    Van Horn, G.; Mac Aodha, O.; Song, Y.; Cui, Y.; Sun, C.; Shepard, A.; Adam, H.; Perona, P.; and Belongie, S. 2018. The inaturalist species classification and detection dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8769--8778

  23. [31]

    K.; and Shah, M

    Vivanco Cepeda, V.; Nayak, G. K.; and Shah, M. 2023. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. Advances in Neural Information Processing Systems, 36: 8690--8701

  24. [32]

    Vo, N.; Jacobs, N.; and Hays, J. 2017. Revisiting im2gps in the deep learning era. In Proceedings of the IEEE international conference on computer vision, 2621--2630

  25. [33]

    Wang, C.; Pan, X.; Pan, Z.; Wang, H.; and Song, Y. 2025. GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains. arXiv preprint arXiv:2505.18700

  26. [34]

    Weyand, T.; Kostrikov, I.; and Philbin, J. 2016. Planet-photo geolocation with convolutional neural networks. In European conference on computer vision, 37--55. Springer

  27. [35]

    Workman, S.; Souvenir, R.; and Jacobs, N. 2015. Wide-area image geolocalization with aerial reference imagery. In Proceedings of the IEEE International Conference on Computer Vision, 3961--3969

  28. [36]

    Xu, S.; Zhang, C.; Fan, L.; Meng, G.; Xiang, S.; and Ye, J. 2024. Addressclip: Empowering vision-language models for city-wide image address localization. In European Conference on Computer Vision, 76--92. Springer

  29. [37]

    Yamaguchi, S.; and Nishida, K. 2025. Explanation Bottleneck Models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 21886--21894

  30. [38]

    Yang, H.; Lu, X.; and Zhu, Y. 2021. Cross-view geo-localization with layer-to-layer transformer. Advances in Neural Information Processing Systems, 34: 29009--29020

  31. [39]

    Yang, Y.; Panagopoulou, A.; Zhou, S.; Jin, D.; Callison-Burch, C.; and Yatskar, M. 2023. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

  32. [40]

    Ye, J.; Lin, H.; Ou, L.; Chen, D.; Wang, Z.; Zhu, Q.; He, C.; and Li, W. 2024. Where am I? Cross-View Geo-localization with Natural Language Descriptions. arXiv preprint arXiv:2412.17007

  33. [41]

    Yeh, C.-K.; Kim, B.; Arik, S.; Li, C.-L.; Pfister, T.; and Ravikumar, P. 2020. On completeness-aware concept-based explanations in deep neural networks. Advances in Neural Information Processing Systems, 33: 20554--20565

  34. [42]

    Yin, W.; Xiao, Z.; Zhou, P.; Yu, S.; Shen, J.; Sonke, J.-J.; and Gavves, E. 2025. Distributional Vision-Language Alignment by Cauchy-Schwarz Divergence. arXiv:2502.17028

  35. [43]

    Yuksekgonul, M.; Wang, M.; and Zou, J. 2022. Post-hoc Concept Bottleneck Models. In The Eleventh International Conference on Learning Representations

  36. [44]

    H.; Zhou, L.; Dai, X.; Yuan, L.; Li, Y.; et al

    Zhong, Y.; Yang, J.; Zhang, P.; Li, C.; Codella, N.; Li, L. H.; Zhou, L.; Dai, X.; Yuan, L.; Li, Y.; et al. 2022. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16793--16803

  37. [45]

    Zhou, Z.; Zhang, J.; Guan, Z.; Hu, M.; Lao, N.; Mu, L.; Li, S.; and Mai, G. 2024. Img2Loc: Revisiting image geolocalization using multi-modality foundation models and image-based retrieval-augmented generation. In Proceedings of the 47th international acm sigir conference on r...

  38. [46]

    Zhu, S.; Yang, L.; Chen, C.; Shah, M.; Shen, X.; and Wang, H. 2023. R2former: Unified retrieval and reranking transformer for place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19370--19380

  39. [47]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  40. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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