Matching a large number of photos in PScan pro

[JanS]

Hi,

We are trying to match 14 UAV flights, totalling 8000+ photos into a limited number of chunks. The photos have a decreased resolution so that we don't run out of RAM memory.

The processing time of the first steps in the photo matching process seems to linearly increase with the number of photo's. The last step, "estimating scene structure" however seems to exponentially increase with the number of photos. We first tried with all 8000 photos but progress got stuck at 0%. A chunk of 2800 photo's is currently running and may eventually finish after 20 hours or so. Chunks of 600-800 work in less than one hour.

Does anyone have an experience with the maximum feasible limit of number of photos?

Note: RAM memory did not seem to be an issue in this step, we didn't get above 85% commit.

Many thanks,

Jan

[Paulov]

Hi Jan.
This is a very interesting "issue". Today I've aquired an educational license and joined more closely to PhotoScan for my university testing stuf. I'm planning to use large photosets too in the future and would be nice if you could share how did you finally manage to  process all the data.

Regards.

Pablo.

[Alexey Pasumansky]

Hello Jan,

Due to parallelized algorithms on Estimating Scene Structure step the progress displayed is not linear, so 20 hours of estimated processing time would be decreased significantly. And this non-linearity also concerns the step from 0% to 1%.
Estimating Scene Structure step is expected to be performed in 6 hours for 8000 image set on K3930 processor, based on our calculations and experience. And about 1-2 hours for 2800 image set for this processing step.

[JanS]

Hi Alexey, Paulov,

Thanks for your comments. This post is from last year, so possibly current Photoscan versions are faster doing the processing. We eventually split everything up in 3 chunks, the largest one with about 4000 photos. I didn't quite remember the processing time, but it was multiple days, mainly due to estimating scene structure. Our PC was a core i7-2600, so not that much slower than your model unless this step also relies on GPU.

After that project, we didn't process any data set close to this size. On typical datasets of 500-1000 photos the geometry reconstruction takes most of the time.