Hi, thanks everyone for continuing to support Celestia. Back in the day I put together an add-on pack of rough asteroid models. Recently I've been working on a set of tools for automatically building texture maps from images sent back from probes. My first target is Ceres. I thought you might like to see my first results from stitching together several Dawn RC3 images.

Once stitching accuracy and normal map generation is better I'll post some files that you can load into Celestia.

Matt

Hi wronkiew,

welcome at CelestialMatters!


I do like this project of yours! How about applying the approach to the few existing hires photos of Pluto and Charon?

Fridger

_________________

Hi Fridger,

My method requires images showing the terrain under different illumination conditions. Unfortunately that's not going to happen for Pluto, because New Horizons was flying out of the Sun. It's possible I might be able to recover pieces of Charon once more data is received.

for removing low frequency shading i like to use Gmic
for the pic of Pluto the other day


( this is what i used for the SFS and blender render in a different post )
viewtopic.php?f=2&t=772#p13319


i normally use "cubeatt" to get a raw image
run a hipass and rerun cubeatt to re-import it into isis

gmic seams to work better than the isis tool "highpass"

as for stitching " noseam" works rather well
http://isis.astrogeology.usgs.gov/Appli ... oseam.html
see the example #1

have you looked at

for height maps ( stereo ) but dose need stereo images
https://github.com/NeoGeographyToolkit/StereoPipeline

and recovering Albedo
https://github.com/NeoGeographyToolkit/PhotometryTK

_________________
"I don't pitch Linux to my friends, I let Microsoft do that for me."
Using OpenSUSE Leap 15 & CentOS 7.6

Thanks John, those links will be very helpful.

I put up some more images to illustrate my process, which works with un-georeferenced probe images.

Image registration and estimate of lighting direction:

Yes, I know the lighting isn't quite correct yet.

First stage output:



The second two encode the lighting and view directions in texture coordinates.

Then the second stage creates a model (albedo plus surface normal) with surface properties that reproduce the probe images. I don't really have any intermediate images that can illustrate the second stage.

There are a few things this process can do that ISIS and ARC's tools can't, in particular it can derive some of the surface albedo without a DEM. However, I'll likely get further by interoperating with those tools than with a fully custom implementation.

I'm stuck at the moment. Here is a part of the normal map that shows the problem:



Medium-scale structures are reconstructed rather well in areas with good probe coverage. Small scale gets a bit noisy due to parallax and other alignment errors, but it's not too bad. The real problem is the tendency of the solver to tilt the surface on large scales, especially between the seams between input images. I expected this to be a problem, where the best-fitting solution isn't necessarily the most realistic one, but I haven't been able to eliminate the effect so far. I've improved the illumination direction finder to more closely match the probe images, much better than in the screenshot I posted earlier. I've tried averaging multiple good surface normal solutions rather than finding the best one, but doing that destroyed the medium-scale structures while doing little to solve the problem. And I have tried rejecting areas with poor illumination diversity, which helped a bit.

I think what I have to do is solve the elevation model rather than the surface normals. That would force it to come up with a solution that would at least be continuous with the unsolved areas. Then I can generate the surface normals from the bump map.