GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization

Worldwide Geo-localization aims to pinpoint the precise location of images taken anywhere on Earth. This task has considerable challenges due to immense variation in geographic landscapes. The image-to-image retrieval-based approaches fail to solve this problem on a global scale as it is not feasible to construct a large gallery of images covering the entire world. Instead, existing approaches divide the globe into discrete geographic cells, transforming the problem into a classification task. However, their performance is limited by the predefined classes and often results in inaccurate localizations when an image's location significantly deviates from its class center. To overcome these limitations, we propose GeoCLIP, a novel CLIP-inspired Image-to-GPS retrieval approach that enforces alignment between the image and its corresponding GPS locations. GeoCLIP's location encoder models the Earth as a continuous function by employing positional encoding through random Fourier features and constructing a hierarchical representation that captures information at varying resolutions to yield a semantically rich high-dimensional feature suitable to use even beyond geo-localization. To the best of our knowledge, this is the first work employing GPS encoding for geo-localization. We demonstrate the efficacy of our method via extensive experiments and ablations on benchmark datasets. We achieve competitive performance with just 20% of training data, highlighting its effectiveness even in limited-data settings. Furthermore, we qualitatively demonstrate geo-localization using a text query by leveraging CLIP backbone of our image encoder. The project webpage is available at: https://vicentevivan.github.io/GeoCLIP

Paper

Similar papers

Peer review

Reviewer SMLq6/10 · confidence 4/52023-06-29

Summary

This paper tackles the image geolocalization problem. Instead of formulating the task as image retrieval or classification over a set of geospatial bins as is typical, the authors frame it as image-to-GPS retrieval by matching image features with location features. The proposed training strategy leverages Random Fourier Features (RFF) to encode geospatial location across multiple scales and align them with features from a CLIP visual encoder, with several training enhancements. The resulting method achieves state of the art performance on three benchmarks, across all scales, and showcases interesting applications in text geolocalization and classification.

Strengths

- Addresses an important and interesting problem. Image localization is a fundamental problem in vision that continues to receive attention in the literature. - The approach has technical novelty and is of the first to formulate the localization task as an image-to-GPS retrieval problem. The key idea is to relate images directly to locations via aligning image features to location embeddings using contrastive learning. Such an approach uses a continuous model of location whereas traditional geolocalization approaches depend on a large reference database of images with known locations to support image-level matching (via image retrieval or classification) and inferring location from the closest match. Instead, the authors propose the concept of a GPS gallery. Though training still requires geotagged images (image + location), the authors show the proposed approach requires less training data. Further, at inference, the GPS gallery can theoretically be any set of locations to query against, instead of being limited by the reference image set. Geospatial location is represented using NeRF-style location embeddings (Random Fourier Features with an MLP). Results show that infusing the location representations with multiple spatial scales improves geolocalization performance. - Evaluation is extensive and compelling. State-of-the-art (SOTA) performance on multiple datasets at different threshold scales (street, city, region, country, and continent). Near SOTA performance on Im2GPS3k training using only 20% of the training data. - Intriguing applications. The results on text geolocalization and geotagged image classification showcase the utility of the location encoding beyond image-based geolocalization.

Weaknesses

- A central idea of this paper is the concept of modeling location continuously. I was disappointed to see that during evaluation, the GPS gallery (the set of possible locations) was limited to locations sampled from the training set (L230). Though I understand why this was done from an evaluation perspective, it significantly hurts the "global localization" theme. At the very least, I would have liked to see an evaluation of how selecting the GPS gallery at inference (choice of sampling distribution, granularity, etc.) impacts performance. This also raises the issue that the location encoder is not truly global, it is limited by where images have been captured on the globe (as training requires image/location pairs). - Some minor details missing. Section 4.2 does not mention which dataset the experiment was performed on. Comparing the numbers from Figure 3 and Table 1a, it can be assumed to have been conducted on Im2GPS3k, but the dataset should still be mentioned to improve clarity. Similarly, section 4.3 does not mention which dataset(s) was/were used for the ablation study. - The application for text query-based geolocalization is interesting, but the results are lacking. It is only demonstrated qualitatively on a continent scale using a single keyword (desert) in the main paper. While this is interesting, it leaves a lot to be desired in terms of understanding the full or even partial capability of the text encoding for tasks such as geolocalization or digital forensics. - Quality of the presentation could be improved. For example, the introduction could provide additional motivation for the underlying problem. In its current form, the introduction mentions that there are a variety of applications for geolocalization, but does not support this claim with any citations or descriptions of how geolocalization benefits them. Instead, the majority of content is related to why the problem is difficult and how existing approaches are inadequate, repeating much of the related work section content (e.g., binning, classification, and scaling methods). The conclusion section, in particular, lacks flow and could benefit from additional editing. Unclear mathematical notation on L138. What is K? Minor grammatical errors (L49, L107, L111, L372, etc.). - Word choice in the conclusion. L353 suggests that this work “solved worldwide geolocalization”. Of course this problem is far from solved!

Questions

My initial rating is weak accept. This paper introduces a method for training a high-quality geospatial location encoder that interacts with, widely used, CLIP features and demonstrates its use for localization and several applications. This has the potential to benefit the wider community. Quantitatively, how does the proposed approach compare in terms of runtime and memory footprint with existing approaches? How does the text encoding perform at smaller scales (region or city)? Can the text encoding benefit the image-based geolocalization?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Limitations addressed in the manuscript.

Reviewer AAT46/10 · confidence 4/52023-07-04

Summary

The paper addresses the problem of geo-locating an image, i.e. estimating the latitude and longitude of the where the image was taken. Previous works treat the problem as a classification problem (of cells distributed across the globe). In contrast the this work proposes to retrieve the GPS coordinates from a collection of possible locations. In order to do so the authors propose a neural network architecture that encodes the (ground truth) GPS location and the image in a joint feature space and train it via a contrastive learning scheme. During the inference the query image is encoded and the GPS coordinate closest in feature space is taken as the predicted location. The main contribution of the paper is the positional encoding of GPS coordinates via random fourier features [21], which are demonstrated to help represent higher frequency information and hence improve localization at higher spatial resolution (Tab. 2). Experimental results demonstrate state-of-the-art retrieval performance on 3 different datasets.

Strengths

**S1** The encoding of GPS coordinates via random fourier features (RFF) is novel. While RFF have been demonstrated to be useful to capture higher frequency properties for 2d image or 3d coordinates [21], e.g. in NeRF, to the best of my knowledge they have not been used to encode lat/lng coordinates for geo-localization. The authors also demonstrate that an equal earth projection (EEP) (which preserves the area covered) is beneficial for encoding 2d coordinates on the globe, especially for higher accuracy localization (< 1km in Tab. 2). **S2** Good ablation studies that demonstrate the usefulness of the chosen solutions, especially Tab. 2 for EEP and RFF, but also the ablations provided in the sup. material in Sec. 3. This helps the reproducibility of the approach. **S3** The place recognition performance improves considerably over state-of-the-art, even though the taken approach is relatively simple. This can be attributed to the ability of the model to learn an understand of spatial information.

Weaknesses

**W1** The usage of the CLIP frozen backbone is not well motivated and evaluated. As such it remains unclear why CLIP features are needed for the place recognition task, rather than a network trained from scratch or fine-tuned on the task. Experimental evaluation demonstrates how the learned feature is able to relate textual information to geographic locations, but the influence of this capability on predicting lat/lng locations for a query image is not evaluated. **W2** The evaluation of the text-to-location encoding in Sec. 4.5 is very limited. The encoded location is used for image classification, but the paper falls too short in outlining the goal and metric of the chosen classification task. As such the evaluation can not stand by it's own but the reader needs to resort to [4, 31] for context. **W3** The size of the gallery (the quantization of possible GPS locations) is tuned per dataset. Instead a single size across all dataset would foster broader applicability of the approach. Sec. 3.1 of the sup. material suggests that this would in fact be possible, e.g. with a size of ~250k. (See also Q2)

Questions

**Q1** We do you keep the CLIP backbone frozen? **Q2** Why is the gallery of GPS coordinates sampled from the *training* set? Is this on the dataset from [9] which is used for training? Did you consider covering the globe, e.g. uniformly? The dependency of the geographic locations on the training set implies that the test set needs to be covered by the training set. This seems to prevent correct out-of-distribution predictions.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The ability to accurately locate any photo on the earth represents a security and privacy issue. E.g., even if a user ops out of GPS tagging their photos, the presented approach enables to infer the user location from their images -- and the location trace of the user over time. This surveillance capability should be stated as negative societal impact or potential harmful consequence of the approach, when deployed in public.

Reviewer UZhc7/10 · confidence 5/52023-07-06

Summary

This paper proposes an image-to-GPS retrieval method, with an image encoder from the pre-trained CLIP and a tailored GPS encoder. The GPS encoder is implemented by using: i) the Equal Earth Projection to mitigate the GPS distortion; ii) Random Fourier transform with multiple MLPs to obtain high-frequency `positional encoding'. Experiments on standard benchmarks demonstrate the effectiveness of the proposed method.

Strengths

1) This paper is well-written and easy to follow; 2) The idea of this paper is interesting and new; 3) The experimental evaluations are strong, particularly the interesting Section 4.4.

Weaknesses

I only have some minor comments: 1) Though it is interesting to adopt the CLIP pre-trained encoders, I would like to the experiments substituting the CLIP image encoder with Swin encoder (as GeoDecoder [5]), or substituting the Swin encoder in [5] with the CLIP image encoder. This would help readers to clearly understand the contribution of the CLIP image encoder. 2) I would like to see an experiment under the following setting: a) Evenly partition the earth into patches, and use the patch center GPS as the database; b) perform the proposed image-to-GPS retrieval and report the performance; c) Repeat a) and b) with coarse-to-fine patch partitions. 3) Please explicitly list the image-encoding time and the image-to-GPS retrieval time; 4) Please tone down the claim of "However, this approach is not practical as it becomes infeasible to construct a gallery containing images covering the entire world." The reason is that we have a world satellite image gallery.

Questions

Please refer to weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer AH575/10 · confidence 4/52023-07-07

Summary

This work learns an image-to-GPS retrieval approach where the image is a query, and the gallery is a collection of GPS coordinates. The authors propose GeoCLIP, which consists of a location encoder, and an image encoder. The retrieval then is based given a new query image against a collection of gps embeddings. Experimental results show that this work outperforms previous SOTA at the Im2GPS3k and GWS15k datasets, across all distance thresholds.

Strengths

* This is a well written paper, with good motivation, writing and experimental section. * The extra results on "Utility of our location encoder beyond geo-localization" provide some more arguments about the effectiveness of their location encoder * The discussion about encoding GPS with an MLP is very interesting and is backed by good experiments. * Hierarchical learning experiments are solid and seem to provide good justification for the design choices.

Weaknesses

* For Figure 3, I have no baseline to compare against. How would other methods compare in terms of how fast the curve degrades? * For the GPS encoding, how come the "city" scale is the one with the biggest rise? Could it be something related to the choice of parameters for the Random Fourier Features? It might make sense if some of the discussion from the supplementary material can be transferred to the main paper here since this is a key part of the method. * For the hierarchical method, how much does the merging of the outputs from the different encoders affect the overall computational needs?

Questions

I have stated my questions above in the weaknesses, so this is what I would like to get answers from the authors on.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors do address this in their work.

Reviewer vH7S6/10 · confidence 4/52023-07-08

Summary

For global localization task for a query photo, the authors propose a CLIP-inspired image-to-GPS retrieval approach (named ‘GeoCLIP’) where we retrieve the GPS coordinates of an unseen query image by matching it with the gallery of GPS coordinates. Regarding location encoder, some new ideas such as Random Fourier Features and Hierarchical Representation were introduced. The experimental results showed the effectiveness of the proposed method by outpeforming the existing baselines.

Strengths

* Matching GPS location embeddings and image embeddings (GeoCLIP) are introduced into an image localization problem. This is novel. * It is very interesting that the location encoder gets inherent alignment with CLIP’s text features, enabling us to map textual descriptions to geographical coordinates. * The proposed location encoding method employing Random Fourier Features is much improved compared with the straightforward method which is the direct use of an individual MLP to encode GPS coordinates into feature vectors. * The experimental results showed that the proposed method outperformed the existing baselines which employs image search methods or grid classification methods.

Weaknesses

* Matching an image embedding with many location embedding (100K (Im2GPS3k) and 500K (GWS15k) coordinates) is needed to estimate a GPS location.

Questions

* This approach is interesting. However, it seems to assume that one location corresponds to one image feature. Is it always true ? For example, in restraurants, there are many kinds of meal menus. In some places, the view might be changed greatly with viewing direction change or even with a small location movement. * When searching the location of a query image, matching it with a collection of GPS coordinates is needed. The needed gallery size of the GPS coordinate seems to be large. How long does it take to estimate the location of one image in case of 100K (Im2GPS3k) and 500K (GWS15k) coordinates.? It seems to need a relatively long time. How about comparison to the baselines regarding search time ? Of course, all search-based methods has this disadvantage as well. However, for image search, many fast search methods has been proposed such as hash-based search. Is it possible to use any fast search methods for the proposed method ? How about search time and computational cost compared with the classification-based methods ?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

A limitation was mentioned regarding image encoding at times of training.

Reviewer SMLq2023-08-14

Thanks for your response. I have read the other reviews and author rebuttals. All reviewers are in agreement and I will be retaining my initial rating. I think the information provided here (e.g., the strategy for constructing the GPS gallery) and across the other responses is important and I encourage the authors to include these results/discussion points as updates to the supplemental material (or in the main document when appropriate).

Authorsrebuttal2023-08-15

We sincerely thank the reviewer for thoroughly reviewing our rebuttal responses and for retaining the initial rating. As suggested, we will appropriately include the results and other important discussions provided in the rebuttal in the final version of the manuscript, either in the main paper or in the supplementary.

Reviewer UZhc2023-08-17

Thanks for the rebuttal and new experiments. In summary, this paper is interesting and I would maintain my initial rating. I would expect authors to include these new experiments in the paper/supplementary, and discuss the benefit of using CLIP. From the rebuttal, I can see that replacing the CLIP encoder with Swin encoder (as GeoDecoder [5]) would result in inferior performance. When referring to Table (a) in section 4.1, the numbers of [5] are 12.8/33.5/45.9/61.0/76.1.

Authorsrebuttal2023-08-18

We sincerely thank the reviewer for taking the time to thoroughly assess our rebuttal responses, appreciating our work, and finding it interesting. It is very encouraging for us to know that the reviewer is maintaining the rating. As suggested by the reviewer, we will make sure to incorporate the new experimental results and a comprehensive discussion on the merits of CLIP into the final manuscript.

Reviewer AAT42023-08-18

I'd like to thank the authors for their detailed in well argumented rebuttal. I have read the other reviews and author rebuttals. I think the authors have addressed the weaknesses raised by the reviewers, especially about the construction of the gallery and it size, well and provided insightful additional experiments. All reviewers agree towards an acceptance of the paper and I will keep my initial rating. Similar to reviewer SMLq I encourage the authors to include the additional results and discussion wrt. to gallery size/construction in the final version of the paper and/or sup. material.

Authorsrebuttal2023-08-21

Thank you for your thoughtful review and for recognizing the efforts we've put into addressing the weaknesses identified by the reviewers. We genuinely appreciate your support and your decision to maintain the initial rating. We are committed to improving the manuscript based on the feedback provided by all reviewers. As suggested, we will certainly include the additional results and discussions concerning the gallery size and construction in the final version of the paper, ensuring that the comments and observations provided are appropriately integrated.

Reviewer vH7S2023-08-21

Thanks for the response. I have read the other reviews and authors' rebuttals. As results, most of the reviewers' concerns seems to have been resolved. All reviewers are in agreement. I would like to keep my initial rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC