Agisoft Metashape

Agisoft Metashape => General => Topic started by: George on September 02, 2013, 09:36:25 AM

Title: Computational power
Post by: George on September 02, 2013, 09:36:25 AM
Hello guys!
I know that this issue has been well discussed so far. However, time runs and there could be some changes.
I still wonder what could be the best solution for Pscan hardware? Say no limitations with resourses (at reasonable level indeed).
In simple terms what is the best computer to use Pscan?
What is the main feature that influences its performance? RAM, graphic card, CPU, hard drive type etc.?
Has anyone implemented a cluster solution?
Appreciate all ideas.
Cheers
Title: Re: Computational power
Post by: airmap3d on September 02, 2013, 12:22:34 PM
Different steps utilise different components of your computer.  A good CPU(s) is always important for the overall running of the program, dual Xeons perhaps.  The GPU is only used in a few of the latter steps but a high end one can significantly reduce processing time and RAM is very dependant on how many images you are trying to process.  The Agisoft Wiki page gives a good guide on RAM requirements and from what I have heard, if you run out of RAM the program will crash so best to over-invest here.  Also, apparently hard drive specs don't really effect the performance of PS although I run SSD's in RAID 1 as they make everyday running of my computer super fast and secure!

I'm sure there will be other people out there that can guide you better but those are my thoughts.

 ;)
Title: Re: Computational power
Post by: Oli63 on September 04, 2013, 02:08:34 AM
I'm not really sure if hard drive's performance doesn't have a lot of influence on total performance, at least in the early phases of alignment.
Currently I am having exactly this problem of buying which kind of machine for an extreme application of PS which produces tens of thousands of 24 MP photos. When I watch the resource monitor in Windows7 (on a i7 machine with 64 GB), especially in "estimating scene structure", which uses up the majority of processing time in alignment, the CPU load is just around 20%, as well as memory. The hard drives are under full stress for an estimated 3% of total time for a few seconds. I usually process between 1000 and 10 000 photos.

I really wonder if there is a chance to make use of GPUs for processing the alignment, which would by far be my favorite, compared to the alternatives of using many XEONs or say the XEON PHI coprocessor.
Title: Re: Computational power
Post by: George on September 04, 2013, 01:32:56 PM
where do you get 24MP images from Oli?
how they perform with Point cloud rendering? Lossless or noisy or ...
Title: Re: Computational power
Post by: Oli63 on September 04, 2013, 03:41:42 PM
From a Sony NEX-7 with their 16 mm lens. Works fine.
Title: Re: Computational power
Post by: George on September 04, 2013, 03:56:10 PM
Was it RAW format?
What applications are you basically after?
Title: Re: Computational power
Post by: mr map on September 04, 2013, 04:23:18 PM
From a Sony NEX-7 with their 16 mm lens. Works fine.

A bit off topic... But I would very much like to know   (if you would care to share  :) ):
@Oli63
Title: Re: Computational power
Post by: Oli63 on September 04, 2013, 07:46:34 PM
- No, only JPG.
- Surveying pipes.

- Sharpness could be better of course but is good enough for aerial imaging. Doesn't make sense to pay a fortune for Zeiss & others.
- The best aperture in terms of abberation, distortion and vignetting is definitely 5.6. There is very good tests about this on http://www.imaging-resource.com, which should always be your first choice when buying cameras and lenses.
- We use it up to ISO 1600 without any problems. This is more than sufficient, as we are getting exposure times around 1/2000 and better. No need for any corrections.
- Low light? Depends on how low. I wouldn't fly in dawn conditions but cloudy skies are no problem at all (in fact, clouds are the best you can have as the light is well distributed and there's no shadows).
- Our triggering is very...well, tricky and we are quite proud of this idea. We opened up the camera (by an expert, of course) and soldered 3 cables to the shutter release. They are brought outside of the camera body where we can focus and release the camera by shortcutting the 3 cables. Works great.

In my opinion, the NEX-7 is still the best choice for aerial imaging. I've read that in autumn Sony will come up with a NEX-9 for around 3000 $ which is said to have a full format sensor. Sounds really cool.
Title: Re: Computational power
Post by: mr map on September 04, 2013, 08:36:20 PM
@Oli63 - thank you for this information - really valuable!

Then I conclude that the 16mm lens is at least sharp enough to bring out the benefit of the 24MP sensor -  making visible difference in the presented image Mosaic.

The trigger solution is creative, if I just knew where to find an expert to open it up an solder the wires. When you trigger, do you connect all three together at once, or do you apply a slight delay between focus and trigger?

This NEX-9 will certainly attract my attention once it is available  8)
/ Tom
Title: Re: Computational power
Post by: Oli63 on September 04, 2013, 10:43:38 PM
At least you get a 2cm GSD from 80m above ground, which is really attractive and the images are fine.

In every country the camera industry must cooperate with a company that can do the maintenance and repair of their equipment. Ask your local Sony/Canon/Nikon etc. representation who is doing this job for them! You can as well ask some professional photographers.

Watch this: http://www.gethypoxic.com/reviews-tests-hacks/camera-hacks-and-mods/78-adding-a-wired-shutter-release-to-the-nex-5.html or this: http://www.gethypoxic.com/reviews-tests-hacks/camera-hacks-and-mods.html

For releasing the shutter you first have to shortcut ground and the focus cable (and keep it). Then shortcut ground and shutter. Very simple.