Satellite Surface Reconstruction

[SBCV]

Satellite Surface Reconstruction

[Kai-46]

@SBCV Thanks for the tips! You could see the roadmaps on the SatelliteSfM readme.

[Kai-46]

@SBCV Yeah, something like that. But I guess I would prefer allow the users to choose their own stereo matching algorithms, e.g., opencv cv.StereoBM_create, or https://github.com/Kai-46/sps-stereo, or any other state-of-the-art stereo matching algorithms? Or, if they want, they could also use off-the-shelf computer vision optical flow algorithms.

[Kai-46]

@SBCV Thanks for the awesome work! And sorry for the ridiculous late reply. I left the satellite image stereo problem for a while, and now am revisiting it and maintaining the newly written repo SatelliteSfM, in order to make satellite image processing more computer-vision-like. There are a couple of things to do in this roadmap. If you are interested in contributing, we could chat over emails.

[SBCV]

@Kai-46 No worries - it is nice to see that you are still working on this topic. Completely abandoning the project would be extremely pity, since it achieves very impressive results.

The reference to SatelliteSfM is super helpful, since I've stumbled over it only recently and that more or less only by accident. I don't know, but maybe you should considering to add a reference also on "https://github.com/Kai-46/VisSatSatelliteStereo" and "https://kai-46.github.io/VisSat/" to increase its visibility.

There are a couple of things to do in this roadmap. If you are interested in contributing, we could chat over emails.

That sounds great but depends on the specific steps in the roadmap. In the long run, I it is reasonable to adapt the SatelliteSurfaceReconstruction to the new api of SatelliteSfM. Is the roadmap somewhere available?

Btw: Have you considered to replace your adapted colmap (which is a full copy of colmap) with a set of git patch files? I have not checked if it is feasible - but if it is, it allows to integrate the latest changes of colmap directly into the colmap adaption.

Here is a simple example:
https://github.com/Kitware/Danesfield/tree/master/patches
https://github.com/Kitware/Danesfield/blob/19e14936dc149aea610819f1f135b981ffa572f4/Dockerfile#L100

[SBCV]

Looks really good!

Regarding the last point:

release code to run stereo matching on rectified stereo pairs, including both classical and deep ones

Do the classical methods refer to something like s2p or micmac?