Mars 2020 Raw Images

Sorted for Content Diversity


A two-day project by Jake Lee using open-source tools
Visit the NASA Mars 2020 Raw Images Gallery

DEMUD algorithm developed by Kiri Wagstaff et al.
at the JPL Machine Learning and Instrument Autonomy Group
All images credit NASA/JPL-Caltech/ASU/CAB
Last updated 2021-04-07 with 10546 images, results updated daily

If you give DEMUD a massive orchard, it will return with one of every fruit.

Because Perseverance takes many images of the same object by design (such as calibration targets), sorting them chronologically in a gallery results in many pages of similar content. To find all of the different objects and content in the gallery, you have to flip through hundreds of pages of thousands of images.

The DEMUD algorithm solves this problem by prioritizing the set of images with the most diverse and novel content. It quickly identifies one of every item and ranks them for you to look at. If you give DEMUD a massive orchard, it will return with one of every fruit. This can be very useful for quickly finding interesting things in a large dataset.

The DEMUD algorithm is also capable of explaining why it thought an item was interesting enough to bring to your attention. It's not implemented here, but check it out below.

Steps

  1. Extract features from each image with ResNet-18's last fully-connected layer prior to softmax
  2. Find interesting images with DEMUD
  3. Display the top 500 in order of prioritization

Resources

  • DEMUD GitHub Repository
  • "Guiding Scientific Discovery with Explanations using DEMUD", Wagstaff et al., AAAI 2013 [paper link]
  • "Unusual ChemCam Targets Disocvered Automatically in Curiosity's First Ninety Sols in Gale Crater, Mars", Wagstaff et al., LPSC 2014 [paper link]
  • "Interpretable Discovery in Large Image Data Sets", Wagstaff and Lee, ICML WHI 2018 [paper link]
  • "Visualizing Image Content to Explain Novel Image Discovery", Lee and Wagstaff, DMKD 2020 [paper site]

Update History

  • First Day: Initial website design, DEMUD results on entire dataset
  • Second Day: DEMUD results on each instrument for filtering

Instrument Filter

This feature allows you to see diverse images from a specific instrument on the Perseverance rover. We re-ran our algorithm on each instrument separately.

Instrument filtered results only use RGB/multi-channel images from each instrument. Most of the grayscale images are the individual channels of those images, so minimal content is lost.

You may notice in filtered results that images seem to get repetitive and similar in latter pages of the gallery. This is simply because we've run out of diverse and interesting images! Finally, some instruments don't have enough images to fill all 20 pages of the gallery, so "N/A" will be shown.

I personally think these are the most interesting results, have fun!