Our Mars rovers have generated tens of gigabytes of data, most of it images. When analyzing these large archives, or when comparing new images to those already acquired, we can benefit from automated tools to assist in prioritizing images so that the most interesting ones appear first. “Interesting” is a subjective judgment, so any automated system should also adapt to individual user interests.
I will describe our experiments with using the DEMUD (Discovery by Eigenbasis Modeling of Uninteresting Data) system to quickly prioritize Mars rover images and guide the discovery of interesting features. By modeling what the user already knows and/or has already seen, DEMUD can focus attention on the unexpected, facilitating new discoveries. DEMUD also accepts user feedback and uses it to influence its next selections. We have found that (1) DEMUD can successfully learn to ignore common features in the environment while remaining sensitive to unusual phenomena with high scientific value and (2) adapting to user feedback leads to higher satisfaction with the images selected by the system. Ultimately, this system could be used onboard future rovers to prioritize data prior to transmission to Earth.
This is joint work with James Bedell, Ravi Kiran, Jim Bell, Rebecca Castano, and Tara Estlin.
Abstract coming soon