Hardware request of information

[Mohabon]

Hello,

we are new PS Pro licensed users and we would like to find an optmized hardware confguration for educational purposes.

Our target is the best cost/performance ratio, so some suggestions are more than welcome.

Our main activities is to process datas in Vietnam acquired with UAV and 16MPixel camera. The number of photos in a chunk should be 900-1000 and the desired resolution of Buld Geometry is ULTRA.

Our main target is to implement a workstation that can complete this task, maybe working for many hours, but without crashing the computer because resources are not sufficient.

Given that Dual XEON is assumed and RAM will be at least 128 Gb ECC, the main doubt is about the GFX card to implement between:
1) NVIDIA GTX 690 4GB.
2) NVIDIA QUADRO 4000 2GB
3) 2xNVIDIA QUADRO 4000 2Gb
4) TESLA CO2075 6GB.

If PS Pro uses CUDA, then probably solution 4 is strongly better, but if it uses only OpenCL, then solution 2-3 are cheaper and good aswell.

Any suggestion and support is more than welcome!

Kind regards,

[Wishgranter]

Hello With 1000 images, 16 Mpix and ULTRA is for now no way with existing hardware, if want process just one mesh.... Agisoft prepare/work on NET solution but it will take some time that is ready for production.


What sort of project it will be just a land (GIS) or cultural heritage or ?  This is most iportaint question now....

[Mohabon]

Hello Wish,

thank for your answer.

The Mapping project is about accurate DEM for mitigation of floods, so it is really important to have accurate Height in the Model, so ULTRA seems to be quite important. I have done some demo/tests with HIGH and results are not bad, but I would like to improve.
This project involves the University of FAO and Vietnam Universities, so the main resources are students.

Now question is: if 1000 Photos at 16Mpix with ULTRA is impossible, what is a reliable number of photos allowing the processing? 300? 500?

About the GFX card, i noticed that the Super computer described in the previous topic uses 2xNVIDIA Quadro. Is there any benefit at using CUDA with TESLA c2075 or is it a waste of money becos not supported?

Kind regards and thanks again!

[Wishgranter]

Hello Mohabon.....

The Tesla/quadro are slower for sure, they have lower clock on GPUs than "gamer" cards.
About the Nvidia stuff, wait a few days, have some very interesant results comparing to AMD/ATI :-)

For the ULTRA, id depend litle on the "model" or better on what is "scanned" For GIS aplication like you do, seems that will need lower requirements, see botton of this page http://www.agisoft.ru/wiki/PhotoScan/Tips_and_Tricks   

With amount of data that can process on ultra, hmm hard to say but think on aprox. 300 images on 256 GB RAM system for now..

But have outlined a possible solution to AGISOFT team for GIS application ( Possibly for other application too ) , that will enable to process even so big projects like yours. But it will take some time for them to implement it.... But Pscan will be capable of doing it in future :-)

 

[Mohabon]

Hello Wish,

so if i understand it right, you suggest to implement a nice GTX690 with 4Gb of RAM instead of TESLA/QUADRO?

My budget will thank you a lot if you confirm!

[Wishgranter]

In simple response yes, but if wait few more day, you will than even more :-)

[Marek]

@Mohabon:
Did you (or anyone) tested Tesla C-2075 in real with PhotoScan?
I have option to buy one from second-hand for 1.2k$ , so any info will be great help to take decision.

[Wishgranter]

ANY of the PRO card are clocked lower than consumer card, so speed will be lower, for 1,2 k can have 3-4 7970 cards - 6-9times faster as one Tesla..... They are clocked lower because of consumption, heat generation and few other stuff.......

[Alexey Pasumansky]

Hello Marek,

We assume that C2075 will not be faster than GTX680 or GTX TITAN, so if you plan to invest in this graphic card only for PhotoScan processing purposes it's not good idea.

[Marek]

Hello Wishgranter and Alexey,
Thanks for reply.
Purpose is only for PhotoScan.

So fot the similar price as C2075 I have offer for 1 x GTX TITAN 6GB or 2 x ASUS Radeon or 3 x AMD  Radeon HD7970. All new with warranty.

What will be better 1xTitan or 2xASUS / AMD 7970?


M.

[Wishgranter]

what sort of stuff you will do ?

if large mesh models 50+ mil polys then is better with TITAN ( 6 GB RAM - all data are in GPU mem - realtime work )

if will do smaller mesh say 10-20 mil triangles, then is beter to have AMD 7970s

if need jsut speed, then 2x 7970 are like 4x TITAN ( aprox. )

but you need a fast CPU too, or the GPUs will not get enought data from CPU ( say 3,5 Ghz+ s ideal )

[Marek]

Main purpose is small ortophoto - 0.5 - 5 km2, photos from UAV with resolution 5cm/px or better, flight level 50-150m above ground, 12-18 MPx.

Photos - every 2-3 seconds during entire flight, so probably 100-1000 photos in one set. Actual - Canon SX260HS 12 MPx - picture jpeg 4000x3000 px)

Goal is speed (possibility of produce ortophotomap and digital elevation model on-site, just after flight).

Application - university lab for training crisis management situations (forest fires, flodings, etc) - delivery actual data for local GIS system used by students.

[Wishgranter]

then are the 7970s best for you........ 

[airmap3d]

Hi Marek and/or Mohabon,

What sort of accuracy are you requiring with your work? xyz?

I do very similar work to what you are intending on doing in here in Australia and can possibly give you a little insight into the process...

From my experience, 'build geometry' isn't even needed to produce an accurate DEM from a UAV of your specifications.  The sparse point cloud over an area of 50ha can give you up to 2-4 million points.  This is more than enough to produce an accurate surface model.

I achieve absolute accuracy's in the order of +-100mm in the vertical (often better) and always better in the xy.  You will need some other point cloud editing software to produce a 'bare earth' model though.

Hope this helps.

[Porly]

The Mapping project is about accurate DEM for mitigation of floods, so it is really important to have accurate Height in the Model, so ULTRA seems to be quite important. I have done some demo/tests with HIGH and results are not bad, but I would like to improve.
This project involves the University of FAO and Vietnam Universities, so the main resources are students.

If you need a accurate DEM the most important thing is a sufficient number of Ground Control Points and a good overlap, my experience is, to get true below 10cm (in height) you need a GCP-grid by every 150-250m. Even if you are going to reach a better (optimizing-)result with less number of GCP's you never know which accuracy you have between these points.

If you are using a UAS-System you can make 2 flights, and change the flight-direction of the second by 90degrees. Like this you can stabilize your block (Alignment-step). You will get more data, but for building geometry you can disable the half data. It is only to get a good alignment result.

If you need a DEM which (always is 2.5 D) you can use Build Geometry on "Height Field"-Mode. Height Field triangulates the estimated points always in 2.5D that causes lower system requirements. 1000 Photos (16Mpix) on ultra should be possible with 128 gb of ram. (screenshot attached)

But the things I mentioned above are more important for the accuracy of a "good" DEM. By the way, even if you build the geometry on high you still have to decimate the model or no??? Aerial captured models are reaching a big number of triangles. Even with 600 Photos I need to decimate the mesh to 40 million. In my opinion it doesnt make lot of sense to build geometry on Ultra (what would give f.e. 500 millions triangles) and decimate it to 50 mio. However, I have no doubt that the model on Ultra would be more detailed compared to High, but for a DEM High should be enough. It is only my point of view
   

Best regards