TOWARDS AUTOMATIC DETECTION OF SHATTER CONES ON PLANETARY ROVER IMAGES: FIRST RESULTS OF A FEASIBILITY STUDY (Invited Presentation)
The Mars-DL project focused on the automatic detection of shatter cones in Martian image scenes. Machine learning methods trained a convolutional neural network (CNN) to detect shatter cones in Mars rover images acting as a “scientific target consultant” (STC). As such images do not exist in reality, we assembled training images using virtual shatter cones, 3D-digitized from real terrestrial specimens provided by the Natural History Museum Vienna. For the CNN training at least a few thousand annotated images are necessary which was solved by image rendering of 3D scenes obtained from MSL Mastcam (Malin et al, 2010) stereo sequences. One challenge was to create realistic looking sceneries containing several artificially placed virtual shatter cones. We tested the preliminary STC version with additional non-annotated artificial images and terrestrial desert scenes from Ethiopia, which contained manually placed (real) shatter cones. These initial tests showed promising results, but also demonstrated the difficulties in achieving a reliable detection rate.
Funding: Austrian ASAP14 Programme.
Ferrière L., Brandstätter F. and Koeberl C. (2013) Finding Shatter Cones in Meteorites from the Natural History Museum Vienna Collection. 76. Met. Soc. Mtg., Abs. #5193.
French B.M. & Koeberl C. (2010) The convincing identification of terrestrial meteorite impact structures. Earth-Sci. Rev. 98, 123–170.
Malin M. C. et al. (2010) The Mars Science Laboratory (MSL) mast-mounted cameras (Mastcams) flight instruments. LPSC. Abs. #1533.