[X] My Assistant

[Bjorn Jonsson]

It works!! For months I've been trying to get my program for generating DEMs from stereo images to work properly. A recent 'rant' of mine can be found here.

Well, to make a long story short my software now at last works:



The only drawback is that the DEM has a somewhat 'contoured' appearance. This shows up in renderings of the DEM, even after I smoothed it considerably using Photoshop. Still this is much better than the infamous horizontal lines I got in my shape from shading experiments. A rendering showing the DEM:



However, 2-3 meters away from the monitor this looks very nice (the thumbnail image also looks nice ). It turns out that the software has managed to detect most of the craters except for the really small ones. If someone finds a way to process the DEM image and remove the 'contours' without excessive loss of details I'd like to know. Please note that the DEM is a 16 bits/pixel image.

Some additional notes on the DEM: The DEM is a section from a 4320x2160 pixel simple cylindrical map and has an altitude range of approximately 16 km. The lowest points (black) are 778.6 km from Rhea's center and the highest ones (white) are 794.5 km from Rhea's center. This is odd (Rhea's dimensions are 768x763x763 km) and is probably due to inaccuracies in the viewing geometry information that came with the PDS files I used. The overall lower elevation of the upper part of the DEM compared to the lower part is probably another manifestation of this.

I used these two images obtained in January 2005:





North is to the right in the images.

Looking at the images it is apparent that the DEM's heavily contoured appearance at upper left is (probably) due to the fact that there wasn't a very big difference in the viewing geometry in that area whereas in the lower part of the DEM the difference is significant.

Taking a section of Steve Albers' map and draping it over the DEM produces interesting results (The crater at ~(315,260) in the DEM is at ~(3440,1160) in Steve's map). Here is a quick-and-dirty animation:

Rhea experimental DEM animation (4.6 MB)

In contrast, without Steve's map things get less realistic:




However, now the lightsource can be anywhere.

Needless to say I'm happy. The only problem is the huge number of DEMs I want to do: All of Saturn's major satellites (everything that can be done - global if at all possible), various sites on Mars etc. This is a lot of work, the DEM of Rhea required 100+ manually measured control points and then took almost 4 hours to compute on a 3 GHz machine. Also the viewing geometry needs to be refined.

[Bjorn Jonsson]

Generating DEMs for MSL the way route maps are generated now would be fun but there are performance issues so I have some doubts. Manually measuring control points also takes a lot of time (I'm using 112 control points for the Rhea DEM) but I may be able to partially automate the process in the future.

More significantly, I have now managed to make some optimizations and changes with the result that my software is at least 5 times faster than it was a week ago but I still really need more speed. As I suspected the 'contoured' appearance of the DEM was because I was computing the DEM to a subpixel accuracy of only 0.25 (4x4 points per pixel). After increasing this to 6x6 things look much better but now take about 8 hours to compute despite the aforementioned optimizations. Increasing this to 8x8 or even 10x10 would be preferable but is not feasible as the DEM would take several days to compute . I obviously need a much faster computer.

The DEM now looks like this:



The contours are less obvious now. Also the DEM is no longer 'tilted' (the upper half of the earlier DEM was at lower altitudes than the lower half). This is because I'm using improved viewing geometry - the subspacecraft_line_sample in the PDS files was off by roughly 30 pixels. Black (0) now represents points 760 km from Rhea's center and white (65535) 768.33 km. This agrees well with Rhea's mean radius of 764.4 km.

Not unexpectedly, the 'contours' are also less obvious in rendered images of the DEM:



Not perfect but much better than the earlier version. Apparently the size of central peaks somehow gets overestimated rather often but overall I'm happy. I even think this is accurate enough to estimate the depth of some of the craters. For example, the crater at (325,270) in the DEM is approximately 4.5 km deep and its central peak 2-2.5 km high.

I'm curious as to one aspect... are you able to pull the camera positions and orientations relative to the satellite neatly out of some database, or are you having to solve for them as part of the optimization? Bundle adjustment, perhaps?

This is one of the things frustrating me - the PDS files contain information on the viewing geometry but as previously indicated, it's not accurate enough. Apparently I can assume the spacecraft position to be correct but the camera pointing needs to be updated - a 30 pixel error in the position of the subspacecraft point in the image is huge.

Now the next thing to do is to generate some additional DEMs, possibly of Tethys' Ithaca Chasma and Odysseus, Enceladus hi-res, Iapetus etc. - the list is endless. Disappointingly, there do not seem to be any nice images to use for generating DEMs of Mimas' Herschel crater. However, after checking the recently released SPICE stuff for the extended mission it seems to me that great images should be possible during the 10,000 km flyby in Cassini's extended mission.

[algorimancer]

I obviously need a much faster computer.

If you happen to have a multi-processor (or dual/quad-core cpu), and you're developing in C++ or Fortran, you might try using OpenMP to parallelize some of the loops. I've found it fairly easy to learn and use, and have had good results with a dual-processor system - now I'm wishing for a quad-core