How to spped up process?

[Dino]

Even with hi-end computer (24Gb Ram, CPU Intel i7-970, GTX 580) Photoscan is struggling with 500 pictures (10mpx) to generate mesh. (High - 6m) Aligning photos took few hours, but to generate model it took 24h and also 24h more to generate mesh. In those stages it uses all amount of available RAM. I also have activated option to use GTX580 in generating mesh stage.

What are my options to spped up process:

- to use resized photos? (50%)
- to mask all images just to leave object of interest
- maybe try to render without using graphic card
- Does location of the photos on the disk plys role?

[Wishgranter]

Hello Dino, as a many people think that config is enough to deal wit so many photos and the very computation hungry procces, for so huge amount of mem you need at least 64GB ram - max in uniprocessor systems today, if need much more then you need a dual or quad socket system for speedy procces with up to 256 GB ram, in march-april will come out new processors and socket from intel AND the new LRDRAM modules with up to 64 GB in single module for "normal" price. best way is the dual socket AMD or INTEL, with 8GB modules adn the 7970 ATI cards. all depend on how much speed/money can put together. the Photogrametry process we can say "raytrace" reality so is a VERY power hungry process......   resizing photos is a possibility but complicated - better to shoot with more thinking before shoot, the masking help a little a GFX card help a LOT in you case as a another CPU and all photos are loaded in RAM for reconstruction so hdd play NO role in this....... good day.. if need help with dualsocket mobo and ram let me know.....

[James]

Something I do to speed up the process is use all photos for alignment, but then disable maybe 50-75% of them before building geometry.

My 'logic' is that the overlap is only needed to calculate camera position and scene structure. Once that is calculated you do not need multiple photos on the same thing in order to build the gemoetry.

However I never worked on such large data sets, and am not familiar with all the algorithms involved so this may not be an appropriate workflow for some applications!

[andy_s]

Something I do to speed up the process is use all photos for alignment, but then disable maybe 50-75% of them before building geometry.

My 'logic' is that the overlap is only needed to calculate camera position and scene structure. Once that is calculated you do not need multiple photos on the same thing in order to build the gemoetry.

However I never worked on such large data sets, and am not familiar with all the algorithms involved so this may not be an appropriate workflow for some applications!

Can anybody comment on whether that logic is correct ?

[Wishgranter]

Andy its correct, but its a hard-work sometimes to decide what images should be used and what not. It depend on how images are overlapped......

[andy_s]

As always Wishgranter - I appreciate your knowledge. 

Sloth is [one of] my prevalent vice  - while I do not under estimate the this might Python automate that activity ?

[My note- The following PhotoScan funtionality can be accessed from Python scripts:
• Open/save/create PhotoScan projects
• Add/remove chunks, cameras, markers
• Add/modify camera calibrations, ground control data, assign geographic projections and coordinates
• Perform processing steps (align photos, build model, texture, decimate model, etc...)
• Export processing results (models, textures, orthophotos, DEMs)
• Access data of generated models, point clouds, images]

[Wishgranter]

If it is something like Fly over a field - area and you clearly see that its a lot of overlap, can disable every 2nd image or this way. Once im get project that was shoot continuesly, every 1 sec one frame,but the overlap was huge. From 800+ images was enought aprox 50-70 images, so im arranged PHOTOS in  EXPLORER ( when importing photos ) , that im have 10 column and select just 2 columns ( one afther another ) so imimported in two parts 160 images, aligned adn then disabled every 2nd image - so reconstruction was processed from 80 images......

But im think too, it could be fully automatised..........

[tincansassoc]

I agree it would be nice to have a feature to select every "x" photos  instead of selecting them every time by hand... i think about this every time I have to select / deselect a bunch of photos...

[James]

There is a very easy way to disable/enable every 'x' photos.

Resize the 'Photos' pane until it is 'x' photos wide, then drag a marquee around a column 'x-1' photos wide from top to bottom, right click and disable.

Sorry just noticed that's what wishgranter already said.... more or less

[Wishgranter]

James, its OK 

[Porly]

Something I do to speed up the process is use all photos for alignment, but then disable maybe 50-75% of them before building geometry.

My 'logic' is that the overlap is only needed to calculate camera position and scene structure. Once that is calculated you do not need multiple photos on the same thing in order to build the gemoetry.

However I never worked on such large data sets, and am not familiar with all the algorithms involved so this may not be an appropriate workflow for some applications!

But it would mean that Photoscan is not working with overdetermination. So that each point is estimated only one time? With higher overlap(all photos used) we have more possibilities which can be adjusted or compensated --> better precision. Isnt it like that???

[andy_s]

But it would mean that Photoscan is not working with overdetermination. So that each point is estimated only one time? With higher overlap(all photos used) we have more possibilities which can be adjusted or compensated --> better precision. Isnt it like that???

Hi Porly, I had to google what "overdetermination" meant 
But now that I have your comment sounds very plausible [to me]. Hope somebody can comment further.

[Wishgranter]

Yes, as mentioned earlier, from 800 phots im took 160, for align, but then disabled-NOT deleted every 2nd image and cleaned the reprojection error. So you use 80 images for depht map generation..... Its wary from project to project...... But its a way to handle it.

When im do my castles from 600-800 images im using aprox 150-250 for depth generation and meshing. But its sometimes complicated to decide what can be used and what not. in GIS its not so problematic im think..... Im preparing few things to clear with Agisoft team for the tutorials..... 

[Porly]

But the disabling step is only to reduce the time of depthmap calculation procedure??? Or do you even reach better results??

In my opinion the calculation with all pictures causes a denser pointcloud (Model). On the other hand, without "overdetermination"(multiple estimation) we dont know how the stereo pairs are chosen. The bigger the distance between stereo-pairs, then better the results. With a short basis between pictures the angel for triangulation is small what causes errors(In object direction). I was always using all pictures because I was sure that photoscan is using more Stereo-pairs to estimate one point...

[Wishgranter]

It seems that YES im get better results, will expain later after better test and confirmation. Pscan can use "classical" stereopairs or multiview process..... Yes up to some degree is better to have more points but afther some degree you get not more better data..... in short.... try generate report if have PRO version and GCPs, there is clear to see the overlap and....