Hello Alexander,
Could you please specify if the same camera setup has no problems earlier, or camera positions have been changed lately?
It seems that there are not enough matching points on some image pairs for depth maps estimation between them. But since the threshold is adaptive, using only "back" cameras provides holeless result.
So generally speaking it is possible to adjust these threshold in the algorithms, but it could result in higher memory consumption and longer processing time, as more depth maps will be estimated.
Therefore could you please check if the quality of mesh reconstruction in the second case is acceptable, because it may be very noisy and even adjusting thresholds will not result in good quality output. So probably refining camera positions and overlap for the models' back area (since the problems commonly appear there for the current camera setup) will be better way.