SUSTech/Mars-Landforms

Landform Retrieval

Paper | Code

Dataset Summary

This dataset is Task 2 of MarsRetrieval, a retrieval-centric benchmark for evaluating vision-language models (VLMs) on Mars geospatial discovery. Task 2 evaluates concept-to-instance generalization for Martian geomorphology. Given a textual geomorphic concept, the model must retrieve its corresponding visual instances from a curated Martian image gallery. The dataset comprises 1,185 carefully curated image patches collected from CTX and HiRISE imagery. The landforms follow a two-level geomorphology taxonomy:

• 7 major genetic classes (e.g., Aeolian, Volcanic and Fluvial processes)
• 48 geomorphic subclasses (e.g., Aeolian Dunes, Dust Devil Tracks, Yardangs)

Task Formulation

We formulate this task as a text-to-image multi-positive retrieval problem:

• A text query describes a geomorphic subclass.
• Multiple image instances in the gallery are considered valid positives.
• The goal is to rank all gallery images by cosine similarity in the embedding space.

Metrics

We report metrics suitable for long-tailed multi-positive retrieval:

• Macro mean Average Precision (mAP)
• nDCG@10
• Hits@10

How to Use

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("SUSTech/Mars-Landforms")

# Access a sample image and its geomorphic label
print(dataset["train"][0]["image"])
print(dataset["train"][0]["label"])

For detailed instructions on the retrieval-centric protocol and official evaluation scripts, please refer to our 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}
}