Mars-Bench: mb-boulder_det

This is the object detection version of the Boulder dataset, designed to localize boulders on the Martian surface using high-resolution orbital imagery from the HiRISE camera. Each image is annotated with manually curated bounding boxes that delineate individual boulders, with annotations created by planetary scientists to ensure high-quality and scientifically accurate labels. The dataset includes predefined train, validation, and test splits. This is one of the smallest datasets in Mars-Bench with only tens of samples. Given the small-object nature of the task, this dataset presents a valuable benchmark for evaluating object detection models in low-data regimes, a setting of growing interest in the computer vision community. To explore all Mars-Bench resources, visit our project website: https://mars-bench.github.io/