Mars-Bench: mb-s5mars

This is a multi-class segmentation dataset developed to enable semantic understanding of Martian surface terrain using imagery captured by the Mastcam camera aboard the Curiosity rover. It contains 8 classes: Bedrock, Hole, Ridge, Rock, Rover, Sand/Soil, Sky, and Track, representing features commonly encountered during rover-based navigation and scientific exploration. Although the original dataset is reported to be annotated by domain experts, in Mars-Bench we further refine it by merging visually ambiguous classes based on expert analysis to reduce label noise and enhance the robustness of model evaluation. To explore all Mars-Bench resources, visit our project website: https://mars-bench.github.io/