TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval

arXiv:2603.24749 · cs.CV · Submitted 2026-03-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval".

Jane: Here's the gist of the paper. Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, location, and time The central challenge is to learn representations that factor out appearance changes driven by time while preserving the underlying geo-location semantics of a place TIGeR supports flexible input configurations (single-modality and multi-modality queries) and uses the same representation to perform (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time–aware retrieval Extensive experiments show that TIGeR consistently outperforms strong baselines and state-of-the-art methods by up to 16% on time-of-year, 8% time-of-day prediction, and 14% in geotime aware retrieval recall Pruning redundant reasoning tokens The paper formalizes the problem as Geo-Time Aware Image Retrieval, which is the task of retrieving an image captured at the same location at a specified target time rather than one chosen only based on visual similarity.

Tom: The core challenge is to learn representations that factor out appearance changes driven by time while preserving the underlying geo-location semantics of a place.

Jane: Existing approaches primarily rank images based on appearance similarity but are trained to be invariant to the time-of-capture.

Tom: Composed retrieval methods modify visual attributes in the image domain but do not guarantee that the retrieved image comes from the same geographic location.

Title and authors: Tom: So we're looking at this paper titled "TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval," and the authors are David G. Shatwell, Sirnam Swetha, and Mubarak Shah from the Institute of Artificial Intelligence at UCF. Jane, what's their main goal here?

Jane: Well, they want to tackle a problem where you need to understand an image not just by where it is taken and what it looks like now, but also when it was taken. They're formalizing this as Geo-Time Aware Image Retrieval because we need to find an image from the same spot at a specific time, not just anything visually similar.

Tom: That sounds really practical for things like digital forensics or urban monitoring, where you need to know exactly what was happening at a certain place and time. So, what's the big idea behind TIGeR itself?

Jane: The core idea is learning representations that can factor out those appearance changes that happen over time while keeping the location information intact. They are proposing this unified framework because existing methods usually treat location, visual data, and time as separate inputs.

Lu: Exactly. They're moving away from just aligning different modalities one by one and instead using a multi-modal transformer with selfattention to let the image, the GPS coordinates, and the time information actually talk to each other directly.

Meng: I wonder how much of that cross-modal reasoning is actually necessary? When we deploy models in production, we need something that's lean. What's their summary saying about what they're trying to achieve with this unified representation?

Tom: They say TIGeR learns a single geotemporal embedding space, and this space is used for three things: geo-localization, time-of-capture prediction, and finally, that geo-time aware retrieval task. It’s supposed to be the one representation you use for everything.

Jane: So instead of having separate models for finding the place and guessing the time separately, they want one model that inherently understands how where and when are related in an image. They're using contrastive learning objectives to make sure these different pieces of information align within that shared feature space.

Lu: That's interesting because they specifically mention departing from dual-encoder methods, like those in SatCLIP or related approaches, which only look at fixed pairs of modalities. TIGeR uses a shared multimodal transformer and joint contrastive learning so the image can attend to the GPS and time information directly, not just be paired with them separately.

Meng: That sounds ambitious for training complexity. How do you think this unified representation actually performs when you try to deploy it? Does it get bogged down by needing too much data for all those interactions?

Tom: The paper shows that TIGeR consistently outperforms strong baselines, achieving up to sixteen percent on time-of-year prediction, eight percent on time-of-day prediction, and fourteen percent in geo-time aware retrieval recall. Those are solid numbers against state of the art methods.

Title and authors: Jane: Those results show that the learned representations are genuinely well aligned across all three modalities—visual, temporal, and geolocation—which is what you need for a robust system. The way they achieve this alignment is key to their success with geo-time aware image retrieval, where they maximize cosine similarity between the query representation and gallery embeddings.

Lu: Their architecture has three stages: an image encoder V(I), a location encoder G(l), and a time encoder T(t). Then, the second stage, this multi-modal transformer F, refines those modality specific embeddings by learning fine grain relationships through selfattention.

Tom: And they show how they optimize this by setting up contrastive losses for pairs like the visual embedding with the location embedding, and also with time embeddings or fused versions of them. It’s a comprehensive training signal designed to enforce that alignment across all pairs.

Jane: When we look at their analysis on time prediction, it's not just about looking at the image; they compare the resulting score against dedicated galleries of location and time embeddings to estimate both where and when something was captured.

Lu: They use an entropy adaptive weighting in their setup, which is important because it ensures that the prior—the existing knowledge about location or time—only has strong influence when it's actually informative, which helps mitigate boundary effects and domain shifts.

Meng: That sounds like a smart way to handle uncertainty. But what about the practical deployment of this time prediction part? If we use their method for ToY and ToD estimation, how sharp are those predictions really in the real world?

Tom: They show sharp, well defined diagonals in both their time of year and time of day matrices, meaning the predictions closely match the ground truth. The interesting caveat is that most errors occur between adjacent temporal classes, which makes sense because it shows where the model is struggling with fine-grained transitions.

Jane: And when we test how robust this framework is to practical constraints, they show that its performance on geo-location distance thresholds, Tgeo in kilometers, stays fairly stable across both datasets. However, it does vary more with the temporal thresholds like time of year or time of day estimations in days and hours.

Lu: They also tested it on compositional image retrieval tasks where the query has both a location and a target time requirement. This is where they showed substantial improvements, achieving gains from plus three point nine to plus fourteen point seven absolute gain over the state of the art baselines on R at one.

Meng: Those gains sound significant for practical applications, but what about running this model? How much computational overhead does this unified framework introduce compared to just running three separate specialized models?

Tom: The training details show they train for ten thousand iterations with a global batch size of one thousand twenty-four. The image encoder uses a pretrained ViT-L/fourteen CLIP encoder which is kept frozen during training and only extracts visual tokens as described in the paper.

Title and authors: Jane: And their embedding space visualization really helps with understanding the concept; they show that the PCA projection of this joint embedding space aligns visual and temporal features on a shared torus like manifold. The image embeddings are green to yellow, and the time embeddings are purple to orange.

Lu: That circular pattern along the outer ring where neighboring months occupy adjacent regions is a very nice visual way to show how they learned the relationships between time categories. It shows that the model has successfully mapped those different temporal concepts into a shared structure.

Meng: What about scaling this up? If we move from 10k iterations to something more practical for production, does the training process remain stable, or do things start breaking at scale?

Tom: The paper notes that they sample at least sixty-four distinct HEALPix cells in their batch construction and avoid repeated month hour combinations to enforce diversity. That diversity sampling is what helps them manage the complexity during training.

Jane: So, to wrap up on the results, TIGeR provides a unified framework for time, images, and geo-location retrieval by using a multi-modal transformer to learn a shared geotemporal embedding space. This allows it to perform geo-localization, time prediction, and that final geo-time aware retrieval task.

Lu: Their main contribution is moving away from separate encoders toward joint contrastive learning where modalities attend to each other directly, which enables them to capture fine grained geotemporal correlations like how a specific location changes across seasons or lighting conditions.

Meng: For someone who just wants to use this for urban monitoring, the implication is that you can ask a complex query—where and when—and get a much more reliable result than if you just used a simple visual search.

Tom: Exactly, because the model is trained to preserve location identity even when there are huge changes in appearance driven by time. It’s about getting that right, not just getting the best visual match.

Jane: So, TIGeR really shows that modeling cross-modal interactions is essential for accurate geo-temporal reasoning, and it enables significantly more reliable compositional retrieval where you need to match both location and time constraints simultaneously.

Lu: The paper also flags that the model's performance is quite sensitive to the temporal thresholds they set for time of year or time of day estimation, which gives us a clear idea about where we might need to tune things based on our specific use case.

Meng: It’s good to know what it doesn't do, though; they admit that the model still struggles when metadata is incomplete or noisy, which is always a real-world issue when you're dealing with messy data streams.

Tom: So that’s the picture on TIGeR: a unified framework trained with contrastive objectives to create a shared space for location, visual appearance, and time. It sets a high bar for how we should think about joint reasoning in these complex geospatial tasks.

Jane: It's definitely something worth keeping an eye on as applications in digital forensics and environmental analysis become more demanding of this kind of multi-faceted understanding.

The paper's summary: Tom: So to wrap up, TIGeR is trying to be this one unified framework that handles time, location, and images all at once.

Jane: Exactly. It takes the problem of finding a picture from the same spot at a specific time and tries to solve it with one cohesive system instead of stitching together three separate tools.

Lu: The core idea is learning this single geotemporal embedding space using a multi-modal transformer that lets the different pieces of information—the image, the location data, and the time data—actually talk to each other through attention mechanisms.

Meng: So you're not just encoding them separately and hoping they line up later? You’re training one thing to understand how where and when are related in a scene.

Tom: Right. And they use contrastive learning objectives, which basically makes sure that the embeddings for different inputs—say, the image and the location—end up close together in this shared space.

Jane: That's the mechanism they use to capture those rich geo-temporal relationships directly instead of aligning things after training. It’s about learning the relationship from scratch within that unified representation.

Lu: The architecture has three parts: an image encoder, a location encoder, and a time encoder. Then there's this middle part, the multi-modal transformer that refines those embeddings by looking at them together via self-attention.

Meng: I like that structure because it keeps the specialization for each modality while forcing them to interact in the fusion stage. But how much computation is this unified approach adding compared to running three separate encoders?

Tom: Well, they say it's training for ten thousand iterations with a batch size of one thousand twenty-four. The image part uses a frozen, pre-trained visual encoder, which saves some training time there.

Jane: They show the results are really strong—up to sixteen percent improvement on time of year prediction and fourteen percent better recall in geo-time aware retrieval compared to previous methods.

Lu: Those numbers suggest that this joint reasoning approach is genuinely better at capturing those subtle appearance changes driven by time while keeping the location semantics solid. It’s a big step toward making these systems more reliable for real-world use.

Meng: If I was deploying this, my main concern would be running it smoothly on a production system and how much data it needs to keep that unified space accurate.

Tom: They address that by using entropy adaptive weighting in their scoring mechanism for time prediction, which means the model only leans heavily on its prior knowledge when it’s actually giving a strong signal. That helps control the uncertainty.

Jane: So they're not just guessing; they’re using a score defined as cosine similarity between the query representation and dedicated galleries of location embeddings to estimate where and when something was captured.

Lu: The visualization of their embedding space is also helpful; they show that visual features and temporal features align on a sort of torus-like manifold, with time categories arranged in a smooth circular pattern.

Meng: That visual mapping really helps you see *why* the model works, which is useful when debugging when things start failing at scale. But what about the limitations? Where does this framework stop working?

Tom: They admit it's pretty sensitive to the temporal thresholds they set for time of year or time of day estimation, meaning tuning those specific parameters matters a lot. Also, it still struggles with noisy or incomplete metadata.

Jane: So while TIGeR is powerful for clean data, its reliability in messy real-world scenarios depends heavily on how well you tune those temporal boundaries and how complete your input data is.

Lu: But the potential here is that this shared space approach opens up much better compositional retrieval, which means you can ask for a location *and* a time simultaneously and get results that actually match both constraints.

Tom: It sounds like they’ve built something that handles the complexity of geospatial reasoning in one go, moving past the limitations of separate models.

Jane: And that’s where it points us next: we need to see how this unified representation holds up when you move beyond just retrieval and into more complex reasoning tasks.

The paper's improvements: Tom: So we're looking at how TIGeR actually improves things—it’s not just about the initial training, but about what it can *do*.

Jane: Right, they show that this unified framework gives us three major improvements: a better way to learn the combined representation, more robust retrieval capability, and much sharper time prediction.

Lu: Specifically for the representation learning, TIGeR moves away from separate encoders and uses that multi-modal transformer to capture those geo-temporal relationships directly within one shared space.

Meng: So this means instead of having three separate systems that might disagree on a location or time, they're forced to learn how they interact in a single structure.

Tom: And the retrieval capability is much better because it can now do geo-time aware retrieval, which means finding an image from the exact same spot at a specific target hour, even if the lighting or season has totally changed.

Jane: That’s significant because it makes the system far more useful for applications like digital forensics where you need that precise temporal context tied to a physical location.

Lu: The time prediction accuracy is also enhanced through an entropy adaptive reranking scheme, which means the model adjusts how much weight it gives to its prior knowledge based on how uncertain it is about its own prediction.

Meng: That sounds like a good way to manage the uncertainty, but what about the practical deployment of these improvements? Does that complexity translate into a faster or more stable system for production use?

Tom: They show that on time of year and day prediction, they get sharp diagonals in their matrices, meaning the predictions are close to the actual truth. The caveat is that most errors happen between adjacent time classes, which shows where the model is struggling with those quick transitions.

Jane: So even though it’s accurate at predicting specific times, you still need to be careful if you're asking for something outside of those sharp boundaries.

Lu: They also tested this framework on compositional retrieval tasks, which means they can handle queries that ask for both a location and a specific time requirement simultaneously.

Meng: That compositional reasoning part sounds powerful, but how does that improve things compared to just searching by location or just searching by time separately?

Tom: They report substantial gains there—from plus three point nine to plus fourteen point seven absolute gain over the state of the art baselines on rank one retrieval. That’s a big jump in reliability for those complex queries.

Jane: This means for a user, they can ask a very specific question—"find me this exact photo from that street taken last Tuesday at three PM"—and TIGeR is much more likely to find it than older methods.

Lu: The visualization of the embedding space confirms this; seeing those features align on that shared manifold shows that the model has successfully learned to map location and time onto a single, understandable structure.

Meng: Still, they're clear about what doesn't work. They flag that the model still struggles when you feed it really noisy or incomplete metadata because the whole system relies on those inputs being somewhat correct.

Tom: So to sum up these improvements: better joint learning for the representation, higher accuracy in predicting time, and much better performance when you need to match both location and time at once.

Jane: It’s a lot of capability packed into one framework, showing how modeling cross-modal interactions is truly essential for getting reliable geo-temporal answers.

Conclusion: Tom: So to wrap up, TIGeR is this unified framework that handles time, location, and images together by learning one shared space for them all.

Jane: Exactly. It’s a system designed to handle complex questions like finding an image from a specific place at a specific time without needing three different models working separately.

Lu: The big implication is moving away from separate encoders toward this joint training where the modalities attend to each other directly, which lets them capture those geo-temporal relationships much more richly.

Meng: From my side, the main change is that we’re getting a single representation that should be more robust than having three independent models that might fight each other over what’s real.

Tom: And the results show it’s hitting high marks on time prediction and retrieval recall, which means this unified approach actually delivers on its promise of being better overall.

Jane: It really shows that modeling these cross-modal interactions is essential if you want to get reliable answers in areas like environmental analysis or urban monitoring where context matters.

Lu: The visualization of the embedding space confirms this; seeing those features align on that shared manifold really demonstrates how the model has successfully mapped location and time onto a single structure.

Meng: I'm still worried about deploying this at scale; if that unified transformer gets too big, it could become computationally expensive for real-time use in a production environment.

Tom: Well, they acknowledge that the training involves ten thousand iterations with a batch size of one thousand twenty-four which is heavy lifting for the compute side.

Jane: It's important to remember that even with those caveats about noise and incomplete data, TIGeR provides a much stronger foundation for complex geo-temporal queries than what we had before.

Lu: The framework’s ability to handle compositional retrieval, where you ask for both location and time at the same time, is what I find most interesting for future representation learning work.

Meng: That compositional reasoning could be useful when we start dealing with more complex operational tasks that require matching multiple constraints simultaneously.

Tom: So that’s the picture on TIGeR: a unified framework trained with contrastive objectives to create a shared space for location, visual appearance, and time.

Jane: It’s definitely something worth keeping an eye on as applications in digital forensics and environmental analysis become more demanding of this kind of multi-faceted understanding.

Lu: And moving forward, the next logical step is to see how we can adapt this unified representation for even more dynamic or real-time scenarios where the inputs change constantly.

Institute of Artificial Intelligence, University of Central Florida

cs.CV

Submitted: 2026-03-25

Updated: 2026-03-28

Comments: Accepted in CVPR 2026

Journal ref: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 23955-23965

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, location…

Key concepts

Geo-Time Aware Image Retrieval
This is the core task: finding an image that matches both a specified geographic location and a target capture time. Unlike simple visual search, this requires reasoning about where the photo was taken and when it was taken simultaneously, making it robust against changes in appearance over time.
Unified Geotemporal Embedding Space
TIGeR learns a single shared feature space that combines information from the image, location, and time encoders. This unified representation allows the model to capture complex relationships between where an object is and when it was captured without needing separate mechanisms for each modality.
Multi-modal Transformer
This architecture uses self-attention to process visual, location, and temporal embeddings together. It takes modality-specific information and aggregates it into a single representation, learning fine-grained interactions between the different types of data to build a richer understanding of the scene.

Terminology

Summary

Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, location, and time The central challenge is to learn representations that factor out appearance changes driven by time while preserving the underlying geo-location semantics of a place TIGeR supports flexible input configurations (single-modality and multi-modality queries) and uses the same representation to perform (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time–aware retrieval Extensive experiments show that TIGeR consistently outperforms strong baselines and state-of-the-art methods by up to 16% on time-of-year, 8% time-of-day prediction, and 14% in geotime aware retrieval recall

Pruning redundant reasoning tokens

The paper formalizes the problem as Geo-Time Aware Image Retrieval, which is the task of retrieving an image captured at the same location at a specified target time rather than one chosen only based on visual similarity. The core challenge is to learn representations that factor out appearance changes driven by time while preserving the underlying geo-location semantics of a place. p. 0

Where the paper says it

Given a query image and a target time, the task is to retrieve an image captured at the same location at the specified time, rather than one chosen only based on visual similarity (Figure 1). The central challenge is to learn representations that factor out appearance changes driven by time while preserving the underlying geo-location semantics of a place.

Existing approaches primarily rank images based on appearance similarity but are trained to be invariant to the time-of-capture. p. 0
Where the paper says it

However, existing approaches do not address this challenge. Standard image retrieval methods primarily rank images based on appearance similarity [3, 10, 45], but are trained to be invariant to the time-of-capture.

Composed retrieval methods modify visual attributes in the image domain but do not guarantee that the retrieved image comes from the same geographic location. p. 0
Where the paper says it

Composed retrieval methods can modify visual attributes (e. , “with snow,” “at night”) [2, 26], but these transformations operate purely in the image domain and do not guarantee that the retrieved image comes from the same geographic location.

Geo-localization models aim to estimate where an image was taken, and prior work extends this to jointly infer where and when an image was captured, yet these models typically encode each modality independently. p. 0
Where the paper says it

Geo-localization models, on the other hand, aim to estimate where an image was taken, and prior work [29, 41] extends this to jointly infer where and when an image was captured. Yet these models typically encode each modality (image, location, time) independently and align only 1 arXiv:2603.

Without an explicit cross-modal fusion mechanism, they struggle to capture rich geo-temporal relationships. p. 1
Where the paper says it

the final embeddings through contrastive learning. Without an explicit cross-modal fusion mechanism, they struggle to capture rich geo-temporal relationships and to model how visual appearance at a fixed location evolves over time.

How the training signal is built

TIGeR proposes a novel approach using a multi-modal transformer that leverages selfattention to jointly process visual, geo-location, and temporal information. One or more modality-specific embeddings are passed through the multi-modal transformer, aggregating information across modalities and producing a unified representation that is then projected into a joint geotemporal embedding space. p. 1

Where the paper says it

To address this challenge, we propose a novel approach using a multi-modal transformer that leverages selfattention to jointly process visual, geo-location, and temporal information. One or more modality-specific embeddings are passed through the multi-modal transformer, aggregating information across modalities and producing a unified representation that is then projected into a joint geotemporal embedding space.

The learned representation serves as a unified geotemporal embedding that supports (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time aware image retrieval. p. 1
Where the paper says it

By allowing modalities to attend to each other, the model learns geo-temporal correlations directly, rather than aligning separately learned embeddings post hoc, thereby enabling robust retrieval even when metadata is incomplete or noisy. The learned representation serves as a unified geotemporal embedding that supports: (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time aware image retrieval.

The model is trained using a combination of contrastive and classification losses with soft metric targets. p. 2
Where the paper says it

This yields six embeddings {v¯, ¯l, ¯t, vl¯ , vt¯ , ¯lt}, all residing in a shared geo-temporal feature space. Losses TIGeR is optimized using a combination of contrastive and classification losses with soft metric targets.

Contrastive objectives encourage embeddings from different input modalities to align within the shared feature space, forming the foundation for all geo-temporal retrieval tasks. p. 2
Where the paper says it

Losses TIGeR is optimized using a combination of contrastive and classification losses with soft metric targets. The contrastive objectives encourage embeddings from different input modalities to align within the shared feature space, forming the foundation for all geo-temporal retrieval tasks.

Where the gains come from

The model architecture comprises three stages: an image encoder V(I), a location encoder G(l), and a time encoder T(t). The second stage employs a multi-modal transformer F to refine the modality-specific embeddings, learning fine-grained relationships between them via selfattention. p. 2

Where the paper says it

Each modality embedding is then linearly projected and layer-normalized. The second stage employs a multi-modal transformer F to refine the modality-specific embeddings, learning fine-grained relationships between them via selfattention.

For each refined embedding X', pooling is performed and then projected into the shared embedding space via learned projections hx(·). The contrastive losses optimize pairs such as (v¯, ¯l), (v¯, ¯t), (v¯, ¯lt), (¯l, vt̄), and (t̄, vl̄) to align unimodal and fused representations. p. 2
Where the paper says it

We employ five contrastive objectives to align all meaningful modality pairs, excluding the individual location and time embeddings ( ¯l, ¯t), which remain uncorrelated. Specifically, we optimize the pairs (v¯, ¯l), (v¯, ¯t), (v¯, ¯lt), ( ¯l, vt¯ ), and (¯t, vl¯ ).

Analysis on geo-time-aware image retrieval

The model is evaluated on geo-time aware image retrieval by encoding the image and time with TIGeR and fusing their embeddings into a unified geo-temporal representation. Retrieval is performed by maximizing cosine similarity between the fused query representation vt¯Q and gallery embeddings v¯G i. p. 5

Where the paper says it

We precompute gallery embeddings v¯ G i = f(I G i ) and, at inference, obtain the fused query representation vt¯ Q = f(I Q, tQ). Retrieval is then performed by maximizing cosine similarity: (I^Q)^* = \argmax _{I^G} \big ( f(I^Q, t^Q)^\top f(I^G) \big ).

TIGeR consistently outperforms prior methods, achieving higher recall at most ranks on both datasets, indicating that the learned representations are well aligned across visual, temporal, and geolocation modalities. p. 5
Where the paper says it

TIGeR consistently outperforms all baselines, achieving higher recall at most ranks on both datasets. This indicates that the learned representations are well aligned across visual, temporal, and geolocation modalities, and that the fused embeddings successfully encode both location and time information rather than over-emphasizing purely semantic visual similarity.

Analysis on time prediction

For time prediction, the objective is to estimate both the geo-location and time-of-capture of a query image by comparing its embedding against dedicated galleries of location and time embeddings. The resulting posterior-like score for each gallery item x i G is defined as (8). p. 6

Where the paper says it

Treat the cosine similarity as a temperature-scaled log-likelihood of a match, and the classifier output as a prior over classes. The resulting posterior-like score for each gallery item x G i is \label {eq:poe-score} \mathrm {score}(x_i^{G}) \;=\; \frac {(\bar {\mathbf {v}}^{Q})^{\top }\mathbf {x}^{G}_i}{\psi } \;+\; \beta (I^{Q})\,\log P\!

The entropy-adaptive weighting in (9) ensures the prior has strong influence only when it is informative, mitigating boundary effects and domain shift. p. 6
Where the paper says it

Equation (8) allows the retrieval term to pull fine-grained neighbors while the prior guides the search toward plausible classes. The entropy-adaptive weighting in (9) ensures the prior has strong influence only when it is informative, mitigating boundary effects and domain shift.

TIGeR exhibits sharp, welldefined diagonals in both ToY and ToD matrices, indicating that predictions closely match the ground truth, with most errors occurring in adjacent temporal classes. p. 7
Where the paper says it

*indicates methods we replicate, closely adhering to protocols outlined by prior work. Method Retrieval TIGeR-test-86k CVT-test ToY ToD ToY ToD Error↓ Error↓ Error↓ Error↓ Zhai et al.

Ablations of the image retrieval thresholds

The performance of TIGeR is shown to be robust to geo-location distance thresholds Tgeo, given in kilometers. The results are more sensitive to time prediction thresholds (TToY, TToD, given in days and hours) which make intuitive sense. p. 7

Where the paper says it

Qualitative results on geo-time aware image retrieval with images from unseen locations. Given a query image I Q and a target time t Q (shown above each column), the goal is to retrieve an image captured at the same location as the query but at the specified target time.

Across both datasets, recall is fairly stable with respect to Tgeo, while it varies more with the temporal thresholds. p. 5
Where the paper says it

TIGeR consistently outperforms all baselines, achieving higher recall at most ranks on both datasets. This indicates that the learned representations are well aligned across visual, temporal, and geolocation modalities, and that the fused embeddings successfully encode both location and time information rather than over-emphasizing purely semantic visual similarity.

Compositional Image Retrieval

The model is further evaluated on the compositional image retrieval task introduced in GT-Loc [29], which tests whether a model can jointly reason about both geographic location and time-of-capture. TIGeR delivers substantial improvements, achieving a +3. p. 1

Where the paper says it

• We curate both a large-scale training dataset and a benchmark dataset with global geographic coverage, broad temporal diversity, and extensive quality filtering, providing the first large-scale testbed tailored to this task. • We show that TIGeR outperforms prior methods, achieving average improvements of 16% on time-of-year prediction, 8% on time-of-day prediction, and up to 14% over state-of-the-art baselines for time-of-capture prediction and geo-time aware retrieval.

9 to +14.7 absolute gain over the SoTA baselines on R@1 (Table 8). p. 5
Where the paper says it

Table 1 reports quantitative results. TIGeR consistently outperforms all baselines, achieving higher recall at most ranks on both datasets.

This shows that modeling cross-modal interactions is essential for accurate geo-temporal reasoning and enables significantly more reliable compositional retrieval. p. 0
Where the paper says it

Extensive experiments show that TIGeR consistently outperforms strong baselines and state-of-the-art methods by up to 16% on time-of-year, 8% time-of-day prediction, and 14% in geotime aware retrieval recall, highlighting the benefits of unified geo-temporal modeling. Introduction Understanding how a location visually appears across different points in time is essential for many real-world applications, including digital forensics, environmental and climate change analysis, cultural heritage preservation, and augmented reality.

Training details

The model is trained for 10k iterations with a global batch size of 1024. The image encoder uses a pretrained ViT-L/14 CLIP encoder [2603. p. 2

Where the paper says it

The first stage includes an image encoder V(I) = V ∈ R NV ×d , a location encoder G(l) = L ∈ R NL×d , and a time encoder T (t) = T ∈ R NT ×d , where NV , NL, and NT denote the number of tokens per modality and d the embedding dimension. The image encoder is a frozen CLIP ViT [24] that outputs one [CLS] token and NV −1 patch embeddings.

24749], which is kept frozen during training and only serves to extract visual tokens as described in Section 4 of the main paper. Location and time are encoded with random Fourier feature (RFF) encoders, each followed by a small MLP. p. 2
Where the paper says it

The image encoder is a frozen CLIP ViT [24] that outputs one [CLS] token and NV −1 patch embeddings. The location and time encoders use random Fourier features (RFF) to project 2D inputs into NL and NT high-dimensional vectors, modulated by frequencies σi ∈ {2 2i} N−1 i=0 .

The fusion transformer is implemented as a single Transformer block that operates jointly on the concatenated tokens from the active modalities. p. 1
Where the paper says it

Without an explicit cross-modal fusion mechanism, they struggle to capture rich geo-temporal relationships and to model how visual appearance at a fixed location evolves over time. To address this challenge, we propose a novel approach using a multi-modal transformer that leverages selfattention to jointly process visual, geo-location, and temporal information.

Batch construction enforces diversity at the batch level by sampling at least 64 distinct HEALPix cells and avoiding repeated month–hour combinations.

Embedding space visualization

The PCA projection of the joint embedding space shows that TIGeR aligns visual and temporal features on a shared, torus-like manifold. Image embeddings are shown with a green–yellow colormap, and time embeddings with a purple–orange colormap. p. 5

Where the paper says it

This indicates that the learned representations are well aligned across visual, temporal, and geolocation modalities, and that the fused embeddings successfully encode both location and time information rather than over-emphasizing purely semantic visual similarity. Qualitative examples comparing TIGeR and the baselines are shown in Figure 5.

Both image and time embeddings trace a smooth circular pattern along the outer ring, with neighboring months occupying adjacent regions. p. 3
Where the paper says it

Benchmark Dataset Curation Constructing a dataset suitable for time-aware geolocalization presents significant challenges. One potential dataset is AMOS [8, 9], which provides a vast collection of static outdoor imagery distributed across diverse regions, months, and times-of-day, however, its raw form is highly unsuitable for direct use.

Conclusion

We introduced TIGeR, a unified framework for jointly modeling images, geo-location, and time. To enable rigorous evaluation, we curated a large-scale benchmark with globally distributed, temporally diverse, and carefully filtered imagery. p. 7

Where the paper says it

Conclusion We introduced TIGeR, a unified framework for jointly modeling images, geo-location, and time, and formalized the task of geo-time aware image retrieval, where the goal is to retrieve an image captured at the same physical location as a query but at a specified target time. To enable rigorous evaluation, we curated a large-scale benchmark with globally distributed, temporally diverse, and carefully filtered imagery.

TIGeR combines modality-specific encoders with a multi-modal transformer to map all modalities into a shared geo-temporal embedding space. p. 7
Where the paper says it

To enable rigorous evaluation, we curated a large-scale benchmark with globally distributed, temporally diverse, and carefully filtered imagery. TIGeR combines modality-specific encoders with a multi-modal transformer to map all modalities into a shared geo-temporal embedding space, and is trained with contrastive objectives and entropy-adaptive reranking to capture fine-grained relationships between where and when a scene is observed.

It is trained with contrastive objectives and entropy-adaptive reranking to capture fine-grained relationships between where and when a scene is observed. p. 7
Where the paper says it

To enable rigorous evaluation, we curated a large-scale benchmark with globally distributed, temporally diverse, and carefully filtered imagery. TIGeR combines modality-specific encoders with a multi-modal transformer to map all modalities into a shared geo-temporal embedding space, and is trained with contrastive objectives and entropy-adaptive reranking to capture fine-grained relationships between where and when a scene is observed.

TIGeR achieves state-of-the-art performance across geo-localization, time-of-capture prediction, and geo-time aware image retrieval. p. 7
Where the paper says it

TIGeR combines modality-specific encoders with a multi-modal transformer to map all modalities into a shared geo-temporal embedding space, and is trained with contrastive objectives and entropy-adaptive reranking to capture fine-grained relationships between where and when a scene is observed. Extensive experiments show that TIGeR achieves state-of-the-art performance across geo-localization, time-of-capture prediction, and geo-time aware image retrieval.

Acknowledgements

Supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. p. 8

Where the paper says it

Acknowledgements Supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.

References

[1] Guillaume Astruc, Nicolas Dufour, Ioannis Siglidis, Constantin Aronssohn, Nacim Bouia, Stephanie Fu, Romain Loiseau, Van Nguyen Nguyen, Charles Raude, Elliot Vincent, et al. Openstreetview-5m: The many roads to global visual geolocation

[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion

[3] Gabriele Berton and Carlo Masone. Megaloc: One retrieval to place them all

[4] Ron Campos, Ashmal Vayani, Parth Parag Kulkarni, Rohit Gupta, Aritra Dutta, and Mubarak Shah. Gaea: A geolocation aware conversational model

[5] Krzysztof M Gorski, Eric Hivon, Anthony J Banday, Benjamin D Wandelt, Frode K Hansen, Mstvos Reinecke, and Matthia Bartelmann. Healpix: A framework for highresolution discretization and fast analysis of data distributed on the sphere

[6] Lukas Haas, Michal Skreta, Silas Alberti, and Chelsea Finn. Pigeon: Predicting image geolocations

[7] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners

[8] Nathan Jacobs, Nathaniel Roman, and Robert Pless. Consistent temporal variations in many outdoor scenes

[9] 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

[10] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition

[11] Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, generalpurpose location embeddings with satellite imagery

[12] Parth Parag Kulkarni, Gaurav Kumar Nayak, and Mubarak Shah. Cityguessr: City-level video geo-localization on a global scale

[13] Martha Larson, Mohammad Soleymani, Guillaume Gravier, Bogdan Ionescu, and Gareth J.F. Jones. The benchmarking initiative for multimedia evaluation: Mediaeval 2016

[14] Xiaopeng Li, Wenyuan Xu, Song Wang, and Xianshan Qu. Are you lying: Validating the time-location of outdoor images

[15] Dongheng Lin, Han Hu, and Jianbo Jiao. What time tells us? p. 8

Where the paper says it

2 [15] Dongheng Lin, Han Hu, and Jianbo Jiao. What time tells us?

an explorative study of time awareness learned from static images

[16] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presenceonly geographical priors for fine-grained image classification

[17] Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. CSP: Self-supervised contrastive spatial pretraining for geospatial-visual representations

[18] Radu Paul Mihail, Scott Workman, Zach Bessinger, and Nathan Jacobs. Sky segmentation in the wild: An empirical study

[19] Eric Muller-Budack, Kader Pustu-Iren, and Ralph Ewerth. Geolocation estimation of photos using a hierarchical model and scene classification

[20] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding

[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma. Dinov2: Learning robust visual features without supervision

[22] Rafael Padilha, Tawfiq Salem, Scott Workman, Fernanda A Andalo, Anderson Rocha, and Nathan Jacobs. Content-aware detection of temporal metadata manipulation

[23] Shraman Pramanick, Ewa M Nowara, Joshua Gleason, Carlos D Castillo, and Rama Chellappa. Where in the world is this image? p. 0

Where the paper says it

edu Abstract Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, location, and time. Beyond standard geo-localization and time-of-capture prediction, these applications increasingly demand more complex capabilities, such as retrieving an image captured at the same location as a query image but at a specified target time.

transformer-based geo-localization in the wild

[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark. Learning transferable visual models from natural language supervision

[25] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching

[26] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval

[27] Tawfiq Salem, Scott Workman, and Nathan Jacobs. Learning a Dynamic Map of Visual Appearance

[28] Tawfiq Salem, Jisoo Hwang, and Rafael Padilha. Timestamp estimation from outdoor scenes

[29] David G. Shatwell, Ishan Rajendrakumar Dave, Sirnam Swetha, and Mubarak Shah. p. 9

Where the paper says it

2, 8 [29] David G. Shatwell, Ishan Rajendrakumar Dave, Sirnam Swetha, and Mubarak Shah.

Gt-loc: Unifying when and where in images through a joint embedding space

[30] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? p. 9

Where the paper says it

1, 2, 4, 6, 7, 8, 3, 5 [30] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at?

joint location and orientation estimation by cross-view matching

[31] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval

[32] Sirnam Swetha, Mamshad Nayeem Rizve, Nina Shvetsova, Hilde Kuehne, and Mubarak Shah. Preserving modality structure improves multi-modal learning

[33] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision

[34] Tsung-Hung Tsai, Wei-Cih Jhou, Wen-Huang Cheng, MinChun 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

[35] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding

[36] Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geolocalization

[37] Nam Vo, Nathan Jacobs, and James Hays. Revisiting im2gps in the deep learning era

[38] Tobias Weyand, Ilya Kostrikov, and James Philbin. Planetphoto geolocation with convolutional neural networks

[39] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval

[40] Angelos Zavras, Dimitrios Michail, Begum Demir, and Ioan- ınis Papoutsis. Mind the modality gap: Towards a remote sensing vision-language model via cross-modal alignment

[41] Menghua Zhai, Tawfiq Salem, Connor Greenwell, Scott Workman, Robert Pless, and Nathan Jacobs. Learning geotemporal image features

[42] Zhongliang Zhou, Jielu Zhang, Zihan Guan, Mengxuan Hu, Ni Lao, Lan Mu, Sheng Li, and Gengchen Mai. Img2loc: Revisiting image geolocalization using multi-modality foundation models and image-based retrieval-augmented generation

[43] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Crossview image geo-localization beyond one-to-one retrieval

[44] Sijie Zhu, Mubarak Shah, and Chen Chen.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unified Geo-Temporal Representation Learning: TIGeR learns a unified geotemporal embedding space through cross-modal attention over image, location, and time inputs. This enables the model to capture rich geo-temporal relationships directly, rather than aligning separately learned embeddings post hoc.

  2. Robust Retrieval Capability: The improved system can perform geo-time aware retrieval, allowing it to retrieve an image captured at the same physical location as a query but at a specified target time, making it robust to substantial seasonal, structural, and lighting changes.

  3. Enhanced Time Prediction Accuracy: TIGeR achieves superior temporal prediction by using an entropy-adaptive reranking scheme that modulates the prior weight based on classifier uncertainty, ensuring the model produces sharp, well-localized temporal distributions for both time-of-year and time-of-day estimations.

  4. Improved Geo-Localization Precision: The system can estimate geo-location by combining continuous retrieval signals with discrete priors, using a score defined as "cosine similarity between its visual embedding and dedicated galleries of location embeddings, then aggregate the scores into 1-month (or 1-hour) bins and apply a softmax to obtain a probability distribution over time-of-year (ToY) or time-of-day (ToD)."

  5. Compositional Reasoning: The improved system can handle compositional retrieval by taking a query consisting of both a location and a target time, retrieving an image that matches both attributes, i.e., an image captured at the same place and at the target time.

Abstract

Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, location, and time. Beyond standard geo-localization and time-of-capture prediction, these applications increasingly demand more complex capabilities, such as retrieving an image captured at the same location as a query image but at a specified target time. We formalize this problem as Geo-Time Aware Image Retrieval and propose TIGeR, a unified framework for Time, Images and Geo-location Retrieval. TIGeR supports flexible input configurations (single-modality and multi-modality queries) and uses the same representation to perform (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time-aware retrieval. By preserving the underlying location identity despite large appearance changes, TIGeR enables retrieval based on where and when a scene was captured, rather than purely on visual similarity. To support this task, we design a multistage data curation pipeline and propose a new diverse dataset of 4.5M paired image-location-time triplets for training and 86k high-quality triplets for evaluation. Extensive experiments show that TIGeR consistently outperforms strong baselines and state-of-the-art methods by up to 16% on time-of-year, 8% time-of-day prediction, and 14% in geo-time aware retrieval recall, highlighting the benefits of unified geo-temporal modeling.

Sources

Related papers