If the approach I suggested above shows some noise in the overlap region when you export the final model, here’s one more step that should help:
- Now that you’ve aligned the two into the same coordinate system and scale, you can export the two separate models from CloudCompare, etc. and create a final, clean isosurface without a visible seam using Stanford’s mesh zippering approach which is ideal in a case like yours: https://github.com/sh0/zipper
If you have more time and want a (potentially) higher quality result, you can go back to the beginning in a different system and set high quality feature recognition. I think you’ll find that if you start with good points, you’ll get good matches and the model that results will be complete.
I know you didn’t have an easy time with Pierre’s OpenMVG, so let me suggest again MVE, which is precompiled for Windows in the link I gave you before, and has nice tutorials and documentation:
https://www.gcc.tu-darmstadt.de/home/proj/mve/
Yes, MVE is sequential in its approach to alignment, however the difference from RC is that you’ll have the chance to provide the initial matching pair for sequential alignment. That could be decisive.
Of course, it may be that your image set cannot be brought into alignment by today’s methods. In that case, all I can say is that you should keep the data handy because new contributions are coming all the time, and what was difficult or impossible one year becomes simpler thereafter.