Forum

Author Topic: Computer for large mapping project  (Read 3296 times)

TXPE

  • Jr. Member
  • **
  • Posts: 54
    • View Profile
Computer for large mapping project
« on: May 25, 2017, 04:27:36 AM »
My largest UAV mapping project to date is 1,500 acres (6 sq km).  I am preparing proposal for up to 12,800 acres (52 sq km) and I know I will need bigger/better PC to process this data.

GSD will be about 3.2 cm/pix using Sony QX1 at 400 ft AGL.  I'm estimating about 16,000 pics.  I assume I'll need to run these in chunks.

I have run into persons in the past who say they had issues using Photoscan for this large of a project, though admittedly, that was at least 2 years ago.  I'm hoping Photoscan is up to the task.

With a budget of $2,500 to $3,000 USD, can anyone make some recommendation for CPU, RAM, and video card? I'm not opposed to building from parts.

Thanks,
Doug

TXPE

  • Jr. Member
  • **
  • Posts: 54
    • View Profile
Re: Computer for large mapping project
« Reply #1 on: May 25, 2017, 11:47:23 PM »
To answer one of my own questions, I'm looking at installing Photoscan on Azure NC24.

I'm hoping there are no issues with installing Photoscan on Azure windows platform.

aloerch

  • Newbie
  • *
  • Posts: 15
    • View Profile
Re: Computer for large mapping project
« Reply #2 on: May 26, 2017, 03:05:25 AM »
Here's my experience:

I've processed UAV and manned imagery using Photoscan. For UAV imagery, it's typically a Sony a6000 (similar in image specs to the qx1) and for manned imagery 2 nikon d810's (one RGB, one NIR). The largest UAV area I've processed has been 2 sq km at 1cm GSD and 1,900 images. The largest manned-aircraft project I've processed in Agisoft was 150 sq km, at 8cm GSD with 14,000 images processed together (7K RGB, 7K NIR). I'm using a Puget Systems Genesis model PC with 128GB Ram (more would be better) and 2 GPU's (eVGA 1080Ti and an eVGA 1080FTW).

I've not had problems with processing these datasets (up to 14,000, 36 Megapixel images) in Agisoft, although I do have recommendations for workflows to smooth the process along.

IF your are using RGB and NIR images, you will want to create your camera model calibrations separately before starting step 1 below, if you're using just RGB, then jump right in.

1. Align ALL of your images in the same chunk if your hardware will support it. Mine supports at least 14,000 images being aligned together.

2. Once the alignment is complete, take care of your ground control, tie-point selections, etc. Optimize the camera positions (if you pre-calibrated RGB and NIR cameras, this step won't affect the camera model calibrations).

3. Split your project into chunks if your project is huge or if you are going to use the "High" quality dense point cloud generation. This step, ie. number of chunks, is dependent on what your hardware can handle. Splitting your images into chunks at this point ensures that your final dense point clouds and meshes/models will remained perfectly aligned to one another. For the 14,000 images project, 7,000 were RGB which I split into 2 chunks, and 7,000 were NIR which I split into an additional 2 chunks, so 4 chunks total.

4. Use batch processing to run your dense point cloud processing on each chunk (not including the original alignment chunk that had all of the images). At this point, in batch processing, you can also set up the classify ground points option, which will perform better on smaller chunks rather than one huge chunk.

5. After the batch processing of dense clouds completes, you can merge the chunks without needing to "align" them as they are already pre-aligned.

6. Proceed to create your DSM, DEM, and Ortho


So, the chunking is pretty much the way to go if your datasets are huge and/or if you want really high density point clouds. As long as the chunks are pre-aligned, there's really no problem. Other applications like Menci APS and 3DFlow Zephyr perform chunking automatically after image alignment in the background (batches) in order to accomplish processing of large datasets on limited hardware. This might be something Agisoft would consider implementing as an option, since chunking is both powerful and clunky. Still, to answer your question, Agisoft has no problem with very large datasets.

$2,000-$3,000 for your budget though might not get you hardware anywhere near what I'm using so.... use a greater number of chunks :-P

TXPE

  • Jr. Member
  • **
  • Posts: 54
    • View Profile
Re: Computer for large mapping project
« Reply #3 on: May 26, 2017, 03:33:54 AM »
Thanks for the detailed info!

Since I posted, the project has grown to 16,000 acres (64 sq km).  I'm expecting in the neighborhood of 25,000 pics.  I also realized $3k wasn't enough.

My IT dept is recommending this:
Azure NC24 has 24 CPU cores, 224GB RAM, and 4 K80 GPUs.
https://azure.microsoft.com/en-us/pricing/details/virtual-machines/windows/
Expand on GPU.

Info on K80: http://www.nvidia.com/object/tesla-k80.html

Thanks,
Doug