Using Chunks workflow...

[WickedAndy]

here's a observation based on my last photoscan test...

1. I have a scene that I am reconstructing using about 60 images, I need to do "ultra" quality.

2. if I put all images into one chunk, alignment is perfect but "ultra" quality fails due to lack of ram, etc..

3. I break scene up into 4-5  image chunks, each chunk reconstructs fine, I then use Align chunks...
alignment completely screws up

4. put all images into once chunk, align photos, duplicate chunk 12 times, remove non used photos per chunk, now build geometry, everything is fine, but now can't merge the chunks because photoscan wants me to align them, even though they are already aligned! so just for the hell of it, I try to run align chunks, and once again about 1/2 the chunks are scrambled... I would like to be able to merge chunks even though it thinks they are not aligned.

also why does photoscan detect points on each image again when doing chunk alignment shouldn't those points been saved in the initial phase when building each chunk? seems a little redundant.

seems like an easy solution would be to allow you to merge chunks without requiring the alignment.... granted you follow the above workflow 

[Alexey Pasumansky]

Hello Andy,

Due to the current state of chunk merging procedure you will not get Ultra High quality model after chunks are aligned and merged.

But still the step "4" could be performed if you reference the original chunk before dublicating it (you'll need to place two markers on the model and set up the reference distance between them - it is not required to define exact distance if you do not need correct model scaling). After that each duplicating of the original chunk will give you another referenced chunk with the exactly the same positioning and scaling. So you will not need to align chunks and could proceed to merging procedure.

[WickedAndy]

Thanks for the tip! yes that does work, allowing me to merge the model..
I did end up just optimizing the amount of images I used for the reconstruction, using more than 50 images I would run out of memory during geometry build..

just some interesting specs..

system
dual hex-core xeon
96gb ram
3x quadro 4000

processed 37 full res images from a canon 5D mk2
ultra high reconstruction
processing time 16hrs
output geometry, 30million polys
texture map size, 20k x 20k

idea for future version, perhaps the ability to breakup a reconstruction volume into multiple builds (tiles)
to make it easier to digest for ultra high detail work.....

[olihar]

Just a little question on the side, I have been wondering regarding using full res images for reconstruction if that is indeed needed, or there to say getting the best results.

Reason I am bringing this up is regarding the bayer sensor as pixel at full resolution are being created, therefore 1/3 of the full resolution should be in theory enough detail for the reconstruction and should get roughly the same amount of accuracy as the full resolution.

This of course does not count the texturing step into this as you would always want to use the full res for the texturing part.

Just a little side thought I have been thinking about ever since I started in photogrammetry, and my fascination by the foveon sensor. (I have not done comparison between the bayer and faveon, but would love to see some tests at full res)

Again just a little side note that could maybe in theory save a lot of processing time.

As well, if you are indeed running out of RAM with 96GB I would take into consideration to get something like this for RAM cache.
http://www.ocztechnology.com/products/solid-state-drives/pci-express/workstation.html

[Wishgranter]

Full res is need if work in thing like cultural heritage - my test on ruined castle - HIGH could not "create" details on windows, mostly just holes
ULTRA - full res is used, every pixel is calculated - all details are as it should to be, and mesh decimate was used afther that, detatails preserved. im preparing a full detailed tutorial(s) on this matter, but still few days on our exhibition........

RAM - best way is using 2P or 4P socket machine with a lot of ram slots - and use the 8GB modules, best on this 2P are intel 18 socket boards
4P are AMD, with interesant price for CPUs and they have 32 slots of mem......

RAM is the most importaint thing, so a even low performance CPUs but with a lot of ram.......