Hello, Alexander,
Thank you for providing sample photos.
The scene you are modeling has a very large depth, while the stereo base is also a little bit large. As a result, there are only few matches between the photos in the near range, while the majority of matches come from the distant objects near the horizon. During alignment a few closer matches are interpreted as the noise, and only the matches from the distant objects are used. This leads to incorrect alignment results and failure of the 3d reconstruction process.
The solution to the problem is masking the distant regions on the photos (near the horizon). After such objects are masked out, the alignment succeeds and the model can be reconstructed.
Please note, that masking is supported only in PhotoScan. So it is not possible to obtain the 3d model for this stereo pair using StereoScan.
With best regards,
Dmitry Semyonov
AgiSoft LLC