Taken from the kaggle repository here.

AI4Mars Dataset

A dataset for terrain classification on Mars, specifically focused on Curiosity (MSL) rover data.

Dataset Structure

The dataset contains high-resolution Mars surface images with corresponding semantic segmentation masks for terrain classification.

Features

• image: Original EDR (Engineering Data Record) images from Mars
• label_mask: Semantic segmentation masks with terrain labels
• rover_mask: Binary masks (1 = rover visible)
• range_mask: Binary distance masks (1 = beyond 30m)
• has_masks: Boolean indicating presence of rover and range masks
• has_labels: Boolean indicating presence of segmentation labels

Labels

Terrain classes are encoded as RGB values in the segmentation masks:

• (0,0,0): Soil
• (1,1,1): Bedrock
• (2,2,2): Sand
• (3,3,3): Big rock
• (255,255,255): No label/null

Data Splits

• Train: Crowdsourced labels with:
  • Minimum 3 labeler agreement
  • 2/3 agreement threshold per pixel
  • 30m distance cutoff
  • Rover regions masked out
• Test: Expert-validated labels with:
  • 100% agreement requirement
  • Three versions available based on labeler agreement thresholds

Image Products

All image products share matching base names with different extensions:

1. EDR Images (.JPG): Raw Mars surface images
2. MXY Files (.png): Rover mask products
3. RNG Files (.png): 30-meter range mask products

Notes

• Current version (0.1) only includes Curiosity (MSL) data
• MER (Mars Exploration Rover) data processing is work in progress
• Range masks are derived from PDS (Planetary Data System) products