System Requirements, Parameters, Accuracy and Errors

[birdofprey]

I need to use Agisoft for small photogrammetric surveys with a Phantom drone... we are talking about approximately 500 shots with the classical resolution for these devices... with a bunch of GCP...

I did a test on a pc leaving the same parameters of the tutorial and the processing times are shocking (4 hours just to align the pictures...)

Quad-core Intel Core i7 CPU, Socket LGA 1150
Motherboard: LGA 1150 with 4 DDR3 slots
RAM: DDR3-1600, 4 x 4 GB for 16 GB total
GPU: Nvidia GeForce GTX 470

I would buy a PC similar to that one I used for testing, just with 32 GB and GTX980 TI... I dont wanna lose money on a PC that cant do the job...

As I do not need accuracy to the millimeter... and not even to the single centimeter... I need to figure out how lowering the parameters in the various phases of the workflow can affect the accuracy of the final product...

[shaav@hotmail.de]

Well, there are some settings that you can change when you open the menu for the alignment and go to the advanced options:
- you can reduce the key point limit and the tie point limit which I wouldn't recommend.
- you can set "pair selection" to disabled which will have very little effect
- you reduce the amount of photos which has enourmous impact
- you need a CPU with more cores and/or faster clockspeed.

If I was in your position I would shoot the photos from a bigger distance and therefore reduce its count to 100-200

[birdofprey]

The Quad-Core Intel Core i7 CPU is already a powerful processor, everything more powerful than this has prohibitive prices. The LGA 1150 is already an high-end motherboard, an LGA 2011-v3 costs too much.

I think that the immediate solution would be to increase the flight altitude (30m to 40m), increase a little bit the camera shutter frequency and split surveys into more flights in order to have about 200 photos.

I do not want to change key point limit and tie point limit (respectively 40000 and 10000) and I want to keep pair selection to reference.

My question is... under all these conditions if instead of using "Accuracy = High" I use "Accuracy = Lowest", how this will affect the final result in terms of errors in the measurements ?

[birdofprey]

Update.

I did some tests limiting to 100 the number of photos. Tests were conducted on a low-performance machine (a single core  3 GHz, 4 GB RAM and no GPU) using all the parameters contained in the tutorial except for Build Dense Cloud in which (to avoid the "Not enough memory") error I used Low and Moderate instead of Medium and Aggressive.

The whole process was successfully completed in a couple of hours.

I deduce that the only parameter that makes the BIG difference is the number of photos to be processed simultaneously.

[Alexey Pasumansky]

Hello birdofprey,

Changing accuracy parameter for the Align Photos procedure doesn't have much effect on the memory consumption. But using Lower accuracy means that for the feature points detection PhotoScan will use the downsampled copies of the input images with the factor 8 by each side (while eon High accuracy the original photos will be used). So the alignment results with such accuracy may be unstable and inaccurate. I don't recommend to go under Medium accuracy.

[birdofprey]

I keep doing tests to develop a procedure really suits my needs.

I realized that I only need to align the photos and then generate a not-georefenced dense point cloud from them. I chose to use the same parameters of the tutorial.... in the Align Photos phase I use High, Reference, 40.000 and 10.000; in the Build Dense Cloud phase I use Medium and Aggressive.

So, I think that by limiting pictures at max 250, a similar configuration could be acceptable:

Quad-core Intel Core i7 CPU, Socket LGA 1150
Motherboard: LGA 1150 with 4 DDR3 slots
RAM: DDR3-1600, for 32 GB total
GPU: Nvidia GeForce GTX 780

Am I making the right choice ?