A personal project to replicate the results of the "Deep Mars" paper by Wagstaff et al. In this project I build a simple forward feed Convolutional Neural Network (CNN) to classify image from the Mars orbital image (HiRISE) labeled data set. Training for 5 epochs with the current model yields about 81% accuracy on the testing data.
- map-proj/: Directory containing individual cropped landmark images
- labels-map-proj.txt: Class labels (ids) for each landmark image
- label_data.py: Python dictionary that maps class ids to semantic names
- deps.txt: Dependencies of this project (that can be pip installled)
- classifier_model.py: The tensorflow model and data cleaning
- Liam Niehus-Staab - niehusst
The HiRISE data used in this project comes from the DOI:
10.5281/zenodo.1048301
Idea for this project and the data originates from the following paper:
Kiri L. Wagstaff, You Lu, Alice Stanboli, Kevin Grimes, Thamme Gowda, and Jordan Padams. "Deep Mars: CNN Classification of Mars Imagery for the PDS Imaging Atlas." Proceedings of the Thirtieth Annual Conference on Innovative Applications of Artificial Intelligence, 2017.