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

David G. Shatwell; Sirnam Swetha; Mubarak Shah

Figure 1: TIGeR unifies time, image, and location understanding, enabling geo-localization, time prediction, and geo-time aware image retrieval: given a query image and a target time, it retrieves an image captured at the same location at the specified time.

Figure 2: Architecture of TIGeR . Given an image

I

, GPS coordinates

l

, and timestamp

t

, modality-specific encoders produce visual (

V

), location (

L

), and temporal (

T

) embeddings. Individual and pair-wise modality tokens are passed through a shared multi-modal transformer

\mathcal{F}(\cdot)

, which employs self-attention to learn geo-temporal interactions. Pooled embeddings are projected into a shared embedding space , where a contrastive loss aligns unimodal and fused representations across all modality pairings (e.g.,

V

–

L

,

V

–

T

,

L

–

T

). In parallel, the classification losses with metric targets on discretized location and time classes encourage the model to learn hierarchical cross-modal representations that vary smoothly across the embedding space.

Figure 3: Benchmark dataset curation pipeline. Starting from the AMOS corpus, we randomly sample 1,255 static cameras with broad global coverage and identify diverse corruption modes. We manually label 4,000 images for quality, train a semi-supervised image quality classifier on frozen visual features, and apply it to 8M unlabeled frames to predict a quality score

P(H|I)

. Using thresholds

T_{H}{=}0.7

and

T_{L}{=}0.4

, images are partitioned into high-, medium-, and low-quality subsets, and low-quality frames are discarded. High-quality images are then used to construct a geographically balanced test set by sampling cameras across

10^{\circ}\times 10^{\circ}

latitude-longitude bins and retaining only cameras with at least 500 frames spanning a year, while the remaining high- and medium-quality images form the training set. The resulting benchmark contains 4.5M training images and 86k test images, with no camera overlap between splits.

Figure 4: Dataset statistics and distributions. Top : Geographical distribution of camera locations across the world in the proposed dataset, showing wide coverage across multiple continents. Left : Time-of-year (month) distribution of the captured images. Right : Time-of-day (hour) distribution of the captured images.

Method	Retrieval	TIGeR-test-86k			CVT
Method	Retrieval	R@1 (%)	R@5 (%)	R@10 (%)	R@1 (%)	R@5 (%)	R@10 (%)
Zhai et al. [ 41 ] *	$It\rightarrow I$	1.97	7.05	11.23	0.45	1.59	2.38
Zhai et al. w/ CLIP [ 41 ] *	$It\rightarrow I$	2.60	8.80	13.70	2.95	8.76	13.31
Time-Loc [ 29 ] *	$It\rightarrow I$	0.38	1.73	3.19	15.23	19.28	21.18
GT-Loc [ 29 ] *	$It\rightarrow I$	0.34	1.56	2.94	16.45	20.44	22.59
TIGeR (Ours)	$It\rightarrow I$	3.51	23.30	37.51	14.55	24.46	29.98

Table 1: Geo-time Aware Image Retrieval on images from unseen locations . Given a query image (

I

) and target time (

t

), the goal is to retrieve an image captured at the same geographic location as the query but at the specified target time . Our model significantly outperforms prior approaches, demonstrating strong geo-temporal alignment and robustness to temporal variation. *indicates methods we replicate, closely adhering to protocols outlined by prior work

Method	Retrieval	TIGeR-test-86k		CVT-test
		ToY	ToD	ToY	ToD
		Error $\downarrow$	Error $\downarrow$	Error $\downarrow$	Error $\downarrow$
Zhai et al. [ 41 ] *	$I\rightarrow t$	68.38	3.97	87.37	3.28
Zhai et al. w/ CLIP [ 41 ] *	$I\rightarrow t$	57.51	3.22	68.95	2.8
Time-Loc [ 29 ] *	$I\rightarrow t$	74.87	4.02	65.10	2.86
GT-Loc [ 29 ] *	$I\rightarrow t$	74.58	3.52	78.95	2.68

Table 2: Time prediction on unseen cameras. *indicates methods we replicate, closely adhering to protocols outlined by prior work.

Method	Retrieval	TIGeR-test-86k		CVT-test
Method	Retrieval	200 km	750 km	200 km	750 km
Zhai et al. [ 41 ] *	$I\rightarrow l$	22.05	46.20	3.50	10.50
Zhai et al. w/ CLIP [ 41 ] *	$I\rightarrow l$	28.58	64.85	23.13	57.74
GeoCLIP [ 36 ]	$I\rightarrow l$	21.58	37.58	43.62	60.77
GT-Loc [ 29 ] *	$I\rightarrow l$	21.07	35.93	42.63	59.49
TIGeR (Ours)	$I\rightarrow l$	48.63	65.61	51.90	66.53

Table 3: Geo-localization accuracy . *indicates methods we replicate, closely adhering to protocols outlined by prior work.

Figure 5: 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. For visualization, the actual time-of-capture for each image is overlaid on top of it. Each row shows top-1 retrieval results from different methods. While prior approaches (Zhai et al. [ 41 ] , GT-Loc [ 29 ] ) often retrieve images captured at incorrect times, our TIGeR model accurately retrieves images corresponding to the target time while maintaining location consistency. Best viewed in color.

Setting	Month	Hour	Geo-location
Setting	Error	Error	Error (km)
Base model	76.77	3.10	668.97
+ multi-modal transformer	57.58	2.98	238.58
+ Classification heads	57.20	2.95	234.53
+ Entropy adaptive reranking	57.20	2.93	203.93

Table 4: Cumulative ablation study on the components of TIGeR .

Figure 6: Distribution of corruption modes in AMOS.

Figure 7: Corruption modes in the AMOS dataset. Examples of typical failure modes in raw AMOS imagery, including sensor or hardware issues (black frames, dead pixels, broken lenses, “no feed” screens), challenging environmental conditions (fog, rain, nighttime), imaging and compression artifacts (blur, glare, noise, color shifts), and content or framing errors (indoor views, floor shots, overlays, or static graphics). These corruptions are frequent and often persistent over time, underscoring the need for the filtering pipeline described in Section 4 .

Figure 8: Sample images that were classified as high, medium or low-quality by our semi-supervised data curation pipeline.

Figure 9: Time prediction probabilities . For an input image

I^{Q}

, we compute cosine similarity between it’s visual embedding

\bar{\textbf{v}}^{Q}

and a gallery of time embeddings

\bar{\textbf{t}}^{G}

, sampled at 1-hour and 1-month intervals. The similarity vector is normalized using the softmax operator, resulting in a predicted time probability distribution. Left : Comparison of the predicted time-of-year probabilities for images at roughly the same time-of-day, but different time-of-years. Right : Comparison of the predicted time-of-day probabilities for images at roughly the same time-of-year, but different time-of-days.

Figure 10: Qualitative geo-localization examples on CVT. Each row shows a query image (left) annotated with its ground-truth GPS coordinate, and the corresponding map of retrieved locations (right). We plot the top-1k candidate GPS coordinates predicted by TIGeR , color-coded by cosine similarity with the image embedding (lighter colors indicate higher similarity). The top-1 prediction is highlighted in yellow and the ground truth is marked with a red

\times

. In all three cases, high-similarity predictions concentrate near the ground truth, indicating accurate localization across diverse scenes and regions.

Table 5: Effect of the shared multimodal fusion transformer. We compare a late-fusion baseline (separate transformers for each modality, combined by averaging) with our shared transformer that performs early fusion of image, location, and time embeddings. We report time-of-year (ToY) and time-of-day (ToD) errors, geo-location error (km), and geo-time-aware image retrieval R@10, averaged over TIGeR-test-86k and CVT (lower is better for errors, higher is better for R@10). The shared fusion transformer yields consistent gains across all metrics.

Image	ToY	ToD	Geo-location	Geo-time
Backbone	Error	Error	Error (km)	R@10
DINOv2	65.02	2.99	563.39	23.20
MAE	73.90	3.26	4139.82	15.05
SigLIP 2	75.88	4.60	1965.55	16.71
CLIP	57.19	2.93	203.78	33.75

Table 6: Effect of the image backbone. The pretrained CLIP [ 24 ] ViT backbone results in stronger overall performance compared to DINOv2 [ 21 ] , Masked Autoencoder [ 7 ] , and SigLIP 2 [ 35 ] . We report time-of-year (ToY) and time-of-day (ToD) errors, geo-location error (km), and geo-time-aware image retrieval R@10, averaged over TIGeR-test-86k and CVT (lower is better for errors, higher is better for R@10).

$\mathbf{T_{geo}}$	$\mathbf{T_{ToY}}$	$\mathbf{T_{ToD}}$	TIGeR-test-86k
$\mathbf{T_{geo}}$	$\mathbf{T_{ToY}}$	$\mathbf{T_{ToD}}$	R@1 (%)	R@5 (%)	R@10 (%)
25.00	15	0.50	1.12	9.10	16.81
25.00	60	2.00	11.48	48.17	64.46
12.50	30	1.00	3.51	23.31	37.51
50.00	30	1.00	3.51	23.31	37.53
25.00	30	1.00	3.51	23.31	37.51

Table 7: Ablations of the image retrieval thresholds on the geo-time-aware image retrieval task. We change the thresholds used to determine if an was retrieved correctly and observe that the predictions are robust to geo-location distance thresholds

\mathbf{T_{geo}}

, given in kilometers. The results are more sensitive to time prediction thresholds (

\mathbf{T_{ToY}}

,

\mathbf{T_{ToD}}

, given in days and hours respectively), which make intuitive sense, given that our choice of temporal thresholds (15-60 days, 0.5-2 hours) cover a larger portion of the total possible range of times compared to the geo-location threshold (12.5-250 km).

Method	Retrieval	TIGeR-test-86k			CVT
Method	Retrieval	R@1 (%)	R@5 (%)	R@10 (%)	R@1 (%)	R@5 (%)	R@10 (%)
Zhai et al. [ 41 ] *	$lt\rightarrow I$	6.63	19.17	26.99	2.14	6.30	9.81
Zhai et al. CLIP [ 41 ] *	$It\rightarrow I$	13.90	33.85	43.71	13.64	33.92	43.92
GT-Loc [ 29 ] *	$lt\rightarrow I$	3.18	12.11	19.08	23.33	49.07	60.86
TIGeR (Ours)	$lt\rightarrow I$	17.84	41.07	51.95	31.61	56.43	65.63

Table 8: Compositional Image Retrieval on Unseen Cameras . Given a query geo-location (

l

) and time (

t

), the task is to retrieve an image that matches the query’s geo-temporal attributes. Our model is trained with a contrastive loss,

\mathcal{L}_{C}(\bar{\mathbf{v}},\bar{\mathbf{lt}})

, explicitly designed to align visual and fused geo-temporal modalities. This enables the model to learn rich interdependencies between images and their spatiotemporal context, allowing it to perform tasks that prior methods struggled to solve accurately. Across both the TIGeR-test-86k and CVT datasets, our approach achieves substantial improvements on all evaluation metrics. *Indicates methods we re-implemented following the protocols detailed in prior work.

Method	TIGeR-Test-86k			CVT
Method	N-R@10	S-R@10	N/S	N-R@10	S-R@10	N/S
Zhai et al.	8.47	14.12	0.60	3.35	1.46	2.29
Zhai et al. CLIP	11.43	16.07	0.71	17.09	9.86	1.73
GT-Loc	2.46	3.47	0.71	27.26	17.92	1.52
TIGeR	32.15	43.13	0.75	33.73	26.46	1.27

Table 9: Hemispheric balance in geo-time-aware image retrieval. We report Recall@10 on the northern (N-R@10) and southern (S-R@10) hemispheres, along with their ratio (N/S), for TIGeR-test-86k and CVT. Values of N/S closer to 1 indicate more balanced performance across hemispheres. TIGeR attains both the highest recalls and the most balanced N/S ratios on both datasets.

Figure 11: Confusion matrices for TIGeR and GT-Loc [ 29 ] on the TIGeR-test-86k and CVT benchmarks. Ground-truth and predicted time-of-year (ToY) and time-of-day (ToD) are discretized into 12 monthly bins and 24 hourly bins, respectively. TIGeR exhibits strong diagonal structure in both ToY and ToD matrices, indicating that predictions closely match the ground truth, with most errors occurring in adjacent temporal classes. This highlights TIGeR ’s substantially superior temporal-prediction performance compared to GT-Loc.

Method	Retrieval	TIGeR-test-86k		CVT-test
		ToY	ToD	ToY	ToD
		Error $\downarrow$	Error $\downarrow$	Error $\downarrow$	Error $\downarrow$
Zhai et al. [ 41 ] *	$I\rightarrow t$	68.38	3.97	87.37	3.28
Zhai et al. w/ CLIP [ 41 ] *	$I\rightarrow t$	57.51	3.22	68.95	2.8
Time-Loc [ 29 ] *	$I\rightarrow t$	74.87	4.02	65.10	2.86
GT-Loc [ 29 ] *	$I\rightarrow t$	74.58	3.52	78.95	2.68

Table 10: Time prediction with additional baselines . We train and test TICL [ 15 ] on our proposed dataset, closely adhering to the training protocol outlined by the authors.

Figure 12: PCA visualization of the TIGeR embedding space. We project 2,000 image–timestamp pairs from the CVT test set onto the first three principal components computed from the time embeddings, and apply the same projection to the corresponding image embeddings. Top: embeddings colored by time-of-year (month). Bottom: embeddings colored by time-of-day (hour). Image embeddings (green–yellow) follow the toroidal structure traced by the time embeddings (purple–orange), showing that TIGeR aligns visual and temporal features on a shared, torus-like manifold.

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

Abstract

Figures & tables

Image Retrieval

TimeSpot: Benchmarking Geo-Temporal Understanding in Vision-Language Models in Real-World Settings

GeoSearch: Augmenting Worldwide Geolocalization with Web-Scale Reverse Image Search and Image Matching

DualGeo: A Dual-View Framework for Worldwide Image Geo-localization

Abstract

Image Retrieval

Explore similar work

TimeSpot: Benchmarking Geo-Temporal Understanding in Vision-Language Models in Real-World Settings

GeoSearch: Augmenting Worldwide Geolocalization with Web-Scale Reverse Image Search and Image Matching

DualGeo: A Dual-View Framework for Worldwide Image Geo-localization