Mars-Bench: mb-crater_binary_seg

The mb-crater_binary_seg dataset focuses on binary crater segmentation using THEMIS imagery, aiming to distinguish crater vs. non-crater regions across the Martian surface. It provides a foundation for crater detection and extraction tasks, supporting research in planetary geomorphology and impact feature recognition. Craters are crucial for understanding the geological history, surface age, and impact processes of planetary bodies. Moreover, classifying and segmenting craters enables researchers to evaluate which morphological features are most informative for scientific analysis. Although craters are often described as bowl-shaped, they are not always perfectly circular. To address this, we provide annotations in elliptical form, marking the first known release of crater segmentation using elliptical geometry, in contrast to the circular annotations commonly used in prior datasets. As the original release contained only metadata, we generated the full image-mask dataset using open-source THEMIS data. Due to missing pixels near the poles, the dataset is geographically limited to within approximately ±30° latitude of the Martian equator. To prevent spatial data leakage, geographically disjoint train, validation, and test splits were created: longitudes −180° – 60° for training, 60° – 120° for testing, and 120° – 180° for validation. To explore all Mars-Bench resources, visit our project website: https://mars-bench.github.io/