Dataset for "Prior-Guided and Level-Aware Progressive Structured Pruning for Martain Terraain Object Detection"

Based on the aforementioned data, we rigorously transform the pixel-level semantic masks into precise object bounding boxes, selecting five functionally critical target classes: Bedrock, Rocks, Tracks, Shadows, and Unknown. To mitigate inherent issues such as local image blurriness and severe class imbalance in the initial annotations, we apply systematic data augmentation (e.g., image rotation) and conduct strict manual verification to refine the bounding box boundaries.

Following augmentation and refinement, the dataset is divided into two distinct subsets: a grayscale image subset (DMER) comprising 3,968 images, and a color image subset (DMSL) containing 16,052 images. Both the DMER and DMSL subsets are randomly partitioned into training and validation sets at a strict 8:2 ratio.