On the use of Chunks and Preselection Source to speed up processing

[FIFER]

Hello,

I have a big dataset of images (>26k) taken with a lawn mower pattern and roughly facing down.

My aim is to speed up the processing time.

I am experimenting on a small set of this dataset to figure out an approach that will considerably speed up the processing time. I am testing 3 approaches:

1) Attach estimated GPS (the dataset is originally non-referenced) position to cameras and use Preselection source in just 1 chunk.
2) Split the non-referenced dataset in different partially overlapping chunks, Align within chunks, then Align and Merge chunks
3) Combine 1 and 2: divide set in chunks and use GPS references within chink to speed up.

Please read the attached pdf which details the output of my tests and provide well described screenshots of point clouds (sorry but would not know how to do otherwise).

As a teaser, here are my

Questions:
1) Is using the Preselection Source with GPS position the only way to speed up Processing?
2) Should the use of chunks help speed up the processing? Or
3) Is the use of chunks meant only to address memory problems but it is inherently longer than single chunk processing?
4) Why point based chunk alignment do not reuse already detected key point and need to redo this step?
5) When aligning chunks, Should a “worst” alignment than that obtained in a single chunk be an expected typical result?
6) Am I using Chunks in the wrong way?



Thank you very much for any explanation.
Cheers
Filippo