SUSTech/Mars-Global-Geolocalization

Global Geo-Localization

**Paper** | **GitHub**

Dataset Summary

This dataset is Task 3 of **MarsRetrieval**, a retrieval-centric benchmark for evaluating vision-language models (VLMs) on Mars geospatial discovery. Task 3 simulates planetary-scale discovery by localizing scientific concepts within the global CTX mosaic, which comprises over 1.4 million CTX tiles.

Ground-truth reference points are compiled from published global scientific catalogues of five Martian landforms:

• —Alluvial Fans
• —Glacier-Like Forms
• —Landslides
• —Pitted Cones
• —Yardangs

Task Formulation

• —Text → Image retrieval
• —Image → Image retrieval

The model retrieves top-K candidate tiles from the global index. A retrieved tile is considered correct if its projected center falls within a spatial tolerance radius (default r = 0.5°) of a ground-truth catalogue coordinate.

Metrics

Given the extreme sparsity of positives, we report:

• —AUPRC (Area Under Precision–Recall Curve)
• —Optimal F1@K\* (best F1 over retrieval depth K)

These metrics quantify planetary-scale distribution estimation rather than simple top-K accuracy.

How to Use

from datasets import load_dataset

ds = load_dataset("SUSTech/Mars-Global-Geolocalization")
print(ds)

For detailed instructions on the retrieval-centric protocol and official evaluation scripts, please refer to the Official Dataset Documentation.

Citation

If you find this useful in your research, please consider citing:

@article{wang2026marsretrieval,
  title={MarsRetrieval: Benchmarking Vision-Language Models for Planetary-Scale Geospatial Retrieval on Mars},
  author={Wang, Shuoyuan and Wang, Yiran and Wei, Hongxin},
  journal={arXiv preprint arXiv:2602.13961},
  year={2026}
}