I am working on the registration engine.
In the next version you can now select the scale on which pattern recognition is done.
For regular (not mosaic) stacks, only scales of 1-5 are needed to get optimal registration.
Internally, I had this fixed to scales of 1 to 8.
For the mosaicing, it will be much more efficient to use smaller scales if you have panels that have little overlap. For your mosaic, you can use scales of 8 to 12.
Simply said. A descriptor scale of 10, means the pattern is only 1/10th of the frame’s dimension. So this descriptor size would be needed if 2 frames to be aligned with each other only have 10% overlap.
Your problem frame only had 7% overlap. For this to be registered robustly, you would need a scale of 12 or 13.
Since I had fixed internally to only use scales of 1-8, this became a problem. You can now tweak this.
If you are using scales higher than 8, use the pentagons for more speed ;-)
I am now working on the actual multiple view mosaic algorithm. The result is better but not yet perfect, so I am trying to see if I can improve more.