sarapohland/lunar-navigation

This dataset was obtained from a high-fidelity lunar environment developed in the Gazebo simulator. The dataset contains 8,661 RGB images (sized 240x320) taken from different regions within the lunar environment. All images are labeled as corresponding to one of eight regions:

1. Smooth terrain on light side (855 images)
2. Smooth terrain on dark side (3,110 images)
3. Bumpy terrain on light side (969 images)
4. Bumpy terrain on dark side (1,140 images) 
5. Inside crater on light side (799 images)
6. Inside crater on dark side (952 images)
7. At the edge of a crater (627 images)
8. At the edge of a hill (209 images)

This labeled dataset has been randomly partitioned into a training set of 6,235 images, a validation set of 1,559 images, and a test set of 867 images.

This dataset also contains a set of out-of-distribution (OOD) data points. The in-distribution data contains images from an uninhabited moon, in which images contain the lunar surface, the sky, and shadows. OOD data contains images from a habited moon, in which images additionally contain astronauts and human-made structures (ladders and habitats). There are 2,045 RGB images (sized 240x320) in the OOD set.

The compressed NumPy file contains the following elements: train_data, train_labels, val_data, val_labels, test_data, test_labels, ood_data, and ood_labels. Sample code for interacting with this dataset is available in our GitHub repository. In particular, src/datasets/custom_dataset.py demonstrates how to set up a PyTorch Dataset and src/datasets/setup_dataloader.py demonstrates how to set up a PyTorch Dataloader.

Each pixel in each of the OOD images is either familiar or unfamiliar to the perception model, depending on whether it corresponds to structures present in the training set or not. These familiarity labels are stored in the ood_labels pickle file. To adjust these labels, you can use the src/datasets/create_labels.py script available in our GitHub repository.