TIGeR: One Framework for Time, Images and Geo Retrieval

TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval (CVPR 2026)

David G. Shatwell, Sirnam Swetha, and Mubarak Shah — Institute of Artificial Intelligence, University of Central Florida.

TIGeR builds one shared geo-temporal embedding that covers images, GPS coordinates, and timestamps. A single model can therefore handle geo-time aware image retrieval, time-of-capture prediction, and geo-localization at once. Given a query image and a target time, TIGeR returns an image of the same physical place taken at the requested time instead of relying on raw visual similarity alone.

The Problem TIGeR Addresses

Many real-world tasks call for reasoning about a place's appearance, its location, and when it was photographed all at once. TIGeR treats this as geo-time aware image retrieval: return an image of the same place at a specified target time.

Why It Matters

The model preserves location identity across major seasonal, lighting, and structural changes, making it useful for digital forensics, environmental monitoring, and long-term visual analysis.

Key Result

TIGeR consistently outperforms strong baselines, with up to 16% better time-of-year prediction, 8% better time-of-day prediction, and 14% better geo-time retrieval recall.

Abstract

Many applications in digital forensics, urban monitoring, and environmental analysis need to reason jointly about visual appearance, location, and time. Beyond standard geo-localization and time-of-capture prediction, these applications increasingly call for more complex capabilities, such as retrieving an image captured at the same location as a query image but at a specified target time. TIGeR is a unified framework for Time, Images and Geo-location Retrieval. It handles flexible input configurations across single-modality and multi-modality queries and relies on one shared representation to perform geo-localization, time-of-capture prediction, and geo-time aware retrieval. To support this setting, the work introduces a multistage curation pipeline and a new benchmark containing 4.5M paired image-location-time triplets for training and 86k high-quality triplets for evaluation. By preserving location identity despite large appearance changes, TIGeR retrieves images based on where and when a scene was captured, not just how it looks.

Method

TIGeR maps image, geo-location, and time into one shared geo-temporal feature space. A frozen CLIP visual encoder generates image tokens, while dedicated location and time encoders represent GPS coordinates and timestamps with random Fourier features. A shared multimodal transformer processes these unimodal and bimodal token sets, learning explicit interactions between where, when, and how a place appears. The refined tokens are pooled and projected into a shared embedding space. TIGeR is trained with cross-modal contrastive objectives so compatible image, location, and time inputs align, while auxiliary classification heads with soft metric targets regularize the space to vary smoothly across nearby places and times.

• Flexible Querying: One representation handles both single-modality and fused queries such as I → t, I → l, and It → I.
• Cross-Modal Fusion: Instead of aligning modalities only at the end, the shared transformer lets image, location, and time tokens attend to each other directly.
• Structured Supervision: Soft targets over discretized location and time classes push embeddings to respect geographic and temporal neighborhood structure.

Dataset and Benchmark Curation

The benchmark is built from AMOS webcam data through a semi-supervised quality filtering pipeline. TIGeR identifies corruption modes, trains an image-quality classifier on labeled examples, discards low-quality frames, and then creates geographically balanced train and test splits with no camera overlap. The final dataset holds 4.5M training images and 86k evaluation images, with wide geographic coverage and strong diversity across months and times of day.

• The benchmark spans diverse locations worldwide and keeps broad month and hour distributions.
• Multistage curation removes corrupted imagery and yields balanced train and benchmark splits.

Main Results

Across geo-time retrieval, time prediction, and geo-localization, TIGeR beats prior geo-temporal baselines on both TIGeR-test-86k and CVT.

Geo-time Retrieval

37.51% R@10 on TIGeR-test-86k. TIGeR hits 37.51% R@10 on TIGeR-test-86k and 29.98% on CVT, giving up to 14% higher geo-time retrieval recall than earlier methods.

Time Prediction (ToY)

48.86 days best ToY error on TIGeR-test-86k with Il → t. TIGeR cuts ToY error to 48.86 on TIGeR-test-86k and 47.00 on CVT, for up to 16% better time-of-year prediction.

Time Prediction (ToD)

3.06 hours best ToD error on TIGeR-test-86k with Il → t. TIGeR brings ToD error down to 3.06 on TIGeR-test-86k and 2.35 on CVT, matching up to 8% better time-of-day prediction.

Qualitative geo-time aware retrieval: TIGeR returns images from the same place at the requested time more reliably than prior approaches.

Geo-time Aware Image Retrieval

With a query image and a target time, TIGeR returns an image from the same location at the requested time. It substantially boosts recall on the new benchmark and stays strong on CVT.

• TIGeR-test-86k: R@1 3.51 · R@5 23.30 · R@10 37.51
• CVT: R@1 14.55 · R@5 24.46 · R@10 29.98

Time Prediction

TIGeR improves both time-of-year and time-of-day prediction, and adding location as auxiliary context brings further gains.

• I → t on TIGeR-test-86k: ToY 51.49 · ToD 3.13
• I → t on CVT: ToY 62.88 · ToD 2.73
• Il → t on TIGeR-test-86k: ToY 48.86 · ToD 3.06
• Il → t on CVT: ToY 47.00 · ToD 2.35

TIGeR yields sharp, well-localized temporal distributions for both time-of-year and time-of-day prediction. Retrieved geo-location candidates cluster near the ground truth across a wide range of scenes.

Geo-localization Accuracy

TIGeR delivers the strongest localization accuracy on both datasets, with particularly large gains on the new TIGeR benchmark.

• I → l on TIGeR-test-86k: 48.63 at 200 km · 65.61 at 750 km
• I → l on CVT: 51.90 at 200 km · 66.53 at 750 km
• It → l on TIGeR-test-86k: 48.34 at 200 km · 64.51 at 750 km
• It → l on CVT: 53.40 at 200 km · 67.15 at 750 km

BibTeX

@inproceedings{shatwell2026tiger, title = {TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval}, author = {Shatwell, David G. and Swetha, Sirnam and Shah, Mubarak}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} }

TIGeR project page — Institute of Artificial Intelligence, University of Central Florida.