Setting up a Amazon Web Server ODM server


#1

Hi all,

I am looking to implement a WebODM server on AWS. I am curious as to what is a recommended minimum specification of server. A typical area size is about 50ha from about 500 images. Anyone here with any setup experience on AWS here?


#2

For 500 images, processed at full resolution, to be safe I would have a machine with at least 64GB. You could get away with less if you process them at a lower resolution (by scaling them). Many factors affect RAM usage and with the ODM codebase changing often, we don’t have a solid benchmark table to make recommendations (but we plan to make one, at some point).


#3

Me too! I am running docker on an i7 iMac with 64gb of ram. In docker advanced settings tab I give docker 6 of 8 cores and 40gb of 64gb total ram. I am running 530 images off of a Phantom 3 Advanced drone. I got a very nice output with these settings. However, about 10% of my imagery, all from around the edges of my ROI, did not get rendered. Should I set Docker to use all 64gb? Might not having done so explain why images did not render? Thank you!


#4

Perhaps too much of the area was cropped. Try passing --crop 0 in your tasks settings, while keeping in mind that you need overlap to generate an orthophoto and it’s perfectly expected that at the edges you will see some cropping. That’s just how the process works and it’s not memory related.


#5

Tried crop 0 - but no luck. Still about 10% of my imagery is not rendering. I have now run ODM 6 times with various settings that ought to mosaic all my imagery but no joy. ODM does fine rendering more than half of the imagery at the edge of my area of interest. But the other half has many missing images. Perhaps this ought to be a separate topic as I don’t want to hijack cwg18.

You can easily see the missing images from this scene shot:

My desired mosaic ought to fill the red polygon and have the buffer you see on the west side especially and on the east somewhat. But on the north and south edge imagery is being left out by ODM.


#6

Try increasing min-num-features to 18000. Forest areas are notoriously difficult to stitch.


#7

I will try that thank you. Though I should tell you that in a previous run I set features to 12,000 or 14,000. I will try 18,000 next. However, it is interesting that the portion of the area where the dropouts occur are in the only non forested section. See this image and look at the south east corner where the imagery is missing - that is a wetland. In this image you can also see my flight paths and I have confirmed that imagery was collected over the entire area.

Thank you for the help!!


#8

After many hours of processing - I got this error:

“Process exited with code 137”

I have now run ODM 6 times and cannot get it to process my entire library of images. I have more areas to collect data for. So, I am afraid I need to look for another solution. Trial and error is turning out to be a trial replete with errors.


#9

Have you calibrated your images? Phantom 3 requires correction before being fed into OpenDroneMap due to barrel distortion of the images.


#10

Thank you Steve. No - we had just flown the day before and we skipped calibration before our mission Phantom 3 mission. Are you suggesting we can correct this via some post processing? If so, a link would be most welcome. I would rather not have to fly the are again but will if that is the best way to get our imagery to process. I need to fill my polygon. Thank you so much - great information from users. Just used MME to process the imagery and it covered about 10 percent LESS area than did ODM. Thanks.

Are there any combination of ODM settings to force it to process insufficiently calibrated imagery?


#11

Yes, we recalibrated all 530 images and reran ODM without any improvement in coverage. I just ran ODM in its default settings wanting to isolate any benefit of the recalibration and the coverage was worse than before calibration, that is to say, there were more missing images including for the first time actually holes in the middle of my region whereas previously the missing images were limited to images on the edge of my region. Not sure what to do next except re-fly the areas that are missing, add them to my library of images and ODM again.


#12

Ok, something else is going on. What percentage of overlap did you fly with?


#13

75%. I have confirmed that indeed all the images for the missing section are in my library. However, when I select and process those last 100 images, the images from our last sortie do not process before or after distortion correction. Here is a side my side showing on the left what is the last image in our campaign and on the right where it out to have rendered into GEP:

Something yes, something else is “wrong” with the images from what was out 5th “mop up” flight (purple track in the above image), of the day so that ODB and MME won’t process them. There are also a few images from our first flight that are similarly dropped from the processing. I guess we will go back and just reply those missing areas. But it’d be nice to know what the underlie is.

Sorry to original poster as this matter has nada to do with AWS whatsoever now.


#14

UPDATE: I just ODMed the last 30 images of the 530 and now they process! See the result here:

It is good to know that they are able to be processed but the mystery continues.