Duration of photo alignment

[tom_marti]

Hi there

Last Friday we took 1800 Pictures from a landfill and are now aligning the photos. But..... the process is running since 70 hours and it will still be running 15 hours.

What did we wrong?

We are using (unfortunately) only an i5 CPU, two Nvidia Quadro K2200, and 32GB RAM.

Does anyone have a similar project? Is this duration "normal"?

The CPU is running at 100%. No idea how the GPUs are used....

[Alexey Pasumansky]

Hello tom_marti,

Which parameters have you used for accuracy, tie/key point limits and preselection?

[tom_marti]

this

Parameter: Accuracy=Medium / Pair preselection=Disabled

Can't look up the key points, the system is running


Update: Keypoint=40'000 and Tiepoints=4'000

[rossnixon]

If there is no longitude/latitude data in the images, that could be a reason for it taking a long time.

I have processed 2129 geo-tagged images with an i5 @ 4.2GHz, 16GB RAM, onboard video. Initial alignment may have taken less than 12 hours (I don't recall exactly). I used 20,000 key points & 4,000 tie points

[stihl]

Using generic preselection will greatly reduce the processing time needed. I'd only advice not to use it when generic mode fails to align the dataset properly.

[tom_marti]

Thanks, next time I will try the generic preselection.

The images have no long/lat data in the EXIF, but I imported the coordinates from a text/log file. So a rough position (+/- 5m) is existing.

@rossnixon: did you use the generic preselection?

[Yoann Courtois]

Hi Tom !

It's normal that it takes so long time if you've disabled the preselection. The soft try to find matches between each pictures and the 1799 others...
If your pictures are good quality ones and give enough overlapping, you firstly should use the generic preselection, which create a first rough alignment in low resolution in order to know which pictures have been taken closely. It will reduce the calculation time a lot, even more if it's 1800 nadiral pictures above a huge area.
Then, it's you have quite accurate position (guess 5m is the best you can expect without expensive geodetic material), you can also use the reference preselection, which choose pictures according to their 3D coordinates. Don't forget to be carfeul about the accuracy you give to your pictures positions !

About another point, even I'm not a specialist, I'm quite sure you cannot make the most of your two quaddro GPU is you only have a I5 CPU. But I deeply hope you didn't forget to enabled OpenCL process with your two GPUs.

Regards

[tom_marti]

Hi Yoann

Our Multicopter (Trimble ZX5) has a built-in GPS which normally get a better accuracy. It depends on the weather and the available GPS-satellites.

to get a needed accuracy of below 5cm we set some tachymetric measured markers in the area.

the OpenCL ist enabled, on both GPUs

[Yoann Courtois]

Hi Tom !

To my mind a built-in GPS will never give enough accuracy to your model, because you'll also need an accurate orientation of the cameras. But it really help the photos alignment !
You should try again to align your set with preselections enabled.
I have handled a set of 1300 photos with a not so different computer (I7 and one K4200), and it took only a few hours to align everything !

[tom_marti]

Thanks, I will change the settings and take a re-run.

[rossnixon]

Thanks, next time I will try the generic preselection.

The images have no long/lat data in the EXIF, but I imported the coordinates from a text/log file. So a rough position (+/- 5m) is existing.

@rossnixon: did you use the generic preselection?

I used 'Reference' & 'Adaptive camera model fitting'.
I recommend you update the EXIF data in your images. I think there are programs that will do that from a data (text?) file.

[SAV]

Use Pair preselection: REFERENCE if you the camera location is known. Th setting will speed up the whole process. Don't forget to adjust the camera accuracy (in your case to 5m) in the REFERENCE SETTINGS (tools icon in the reference pane).

Additionally you might want to consider to invest in an i7 quadcore processor because feature detection and image matching which produce the sparse point cloud (tie points) are run on your CPU. More/faster cores, more speed.

Regards
SAV

[tom_marti]

Thanks for all your help.

I was able to re-run the photo alignment with the given parameters.

pair preselection: reference
Key/Tie-points adjusted to 20'000 / 4'000
CPU: i7
GPU: 1 GeForce GTX 680

Now the duration is only 3 hours! 
And the result is perfect