Mars novelty detection Mastcam labeled dataset

These datasets were used for experiments in the paper:
Kerner, H. R., Wellington, D. F., Wagstaff, K. L., Bell III, J. F., Ben Amor, H. Novelty Detection for Multispectral Images with Application to Planetary Exploration. In Proceedings of Innovative Applications in Artificial Intelligence (IAAI/AAAI), 2019. 

The file cae_train.zip contains examples used for training the convolutional autoencoder. Each example is a Numpy array (.npy) of size 64x64x6 pixels.

The file novel_images.zip contains the 332 tiles labeled as "novel" for containing novel geologic features. Each example is a Numpy array (.npy) of size 64x64x6 pixels.

The file cnn_direct.zip contains examples used for fine-tuning pre-trained networks. Images are divided into "vis" (shorter wavelengths) and "nir" (longer wavelengths) and by their label of "typical" vs. "novel." These three-channel images are stored as .jpg files.

All files are named with the following convention: sequence_id_XX*_{R,L}Y_solZZZZ_N.npy where XX* is the sequence ID for the image, {R,L}Y indicates the right (R) or left (L) eye of the camera and the image number in the sequence (Y), and ZZZZ is the four-digit sol (Martian day since the rover began operations) the image was acquired on.

All source images are publicly released Experiment Data Records (EDRs) archived by the Planetary Data System (PDS).