Mars Rover Terrain Segmentation Labels from Crowdsourcing

AI4MARS provides ~326,000 semantic segmentation labels for 35,000 images from NASA's Curiosity, Opportunity, and Spirit rovers. The dataset was created for training terrain classification models to support autonomous driving on Mars, with labels collected via crowdsourcing and validated by NASA mission scientists. Each crowdsourced image label was produced by 10 annotators to ensure quality.

Use Cases

• Train semantic segmentation models on ~326K full image labels to classify Martian terrain types for rover navigation
• Validate model predictions against the ~1.5K expert-annotated labels from NASA rover planners and scientists
• Analyze label agreement and quality across the 10 annotators per image provided through crowdsourcing
• Develop terrain-aware autonomous driving algorithms using labeled imagery from the Curiosity, Opportunity, and Spirit rovers

Strengths

• Large scale of ~326,000 semantic segmentation labels across 35,000 images
• Includes ~1,500 validation labels annotated by expert NASA rover planners and scientists
• Each crowdsourced image label was reviewed by 10 annotators to improve quality and agreement
• Data originates from NASA's MSL and MER missions, providing authoritative Martian imagery

Limitations

• Dataset composition and potential class balance are not described
• The raw description does not specify the resolution, spectral bands, or preprocessing of the 35K source images
• No information provided on the specific terrain classes or label taxonomy used in the segmentation

Provenance

Source

National Aeronautics and Space Administration (NASA)

Collection Method

Labels collected through crowdsourcing, with a validation set annotated by NASA rover planners and scientists from the MSL and MER missions.

Freshness

Last updated on 2025-08-22.

Geography

Mars, from the Curiosity, Opportunity, and Spirit rover missions.