Forum

Author Topic: Image Collection Limits  (Read 25384 times)

Nathan Craig

  • Newbie
  • *
  • Posts: 18
    • View Profile
    • Personal Page
Image Collection Limits
« on: July 22, 2010, 07:00:44 AM »
From a thread on a kite aerial photography (KAP) discussion board, I just learned of PhotoScan (http://ostro.ced.berkeley.edu/~crisr/discuss/comments.php?DiscussionID=1595&page=3). I am an archaeologist using KAP to photograph and document large sites. I have been working with several structure from motion (SfM)packages including Photosynth, Bundler, and Photocity. Today, I downloaded the PhotoScan demo and was extremely impressed. It looks like PhotoScan will serve my needs very well, and I just purchased an educational license of the PhotoScan.

For recording archaeological sites with KAP, I focus on straight down "ortho" photographs. I generally use a gyro-stabilized rig, and I freuqntly produce planar panoramas consisting of over a hundred images. Many of the sites in Peru that I am recording with KAP are greater than 10 ha. Thus, my requirements for SfM involves processing image collections of several hundred photographs. Using PhotoScan, I have processed collections of over 40 images but have not yet attempted a collection of more than a hundred images.

I am curious as to what are the limiting factors on the size of image collections that can be processed. By this I mean how many images can I expect to process, and what are the specific limiting factors for the number of images that can be processed? Are there hardware configurations that would facilitate or permit the processing of more images? Are there software or operating system "caps" that prevent processing collections of more than some number of images or perhaps better stated as total number of pixels?

Based on my initial tests, it appears that 4 megapixel images produce very good results. With a hypothetical size of 4 megapixel images, how might I go about calculating the maximum size of image collections that I can process. My interests are in using a machine that will maximize the number of images that can be processed. Any insight into these issues would be greatly appreciated.

Thank you,
Nathan

Dmitry Semyonov

  • Agisoft Technical Support
  • Full Member
  • *****
  • Posts: 200
    • View Profile
Re: Image Collection Limits
« Reply #1 on: July 23, 2010, 07:58:52 PM »
Thank you for purchasing AgiSoft PhotoScan.

PhotoScan doesn't explicitly impose any limits on the number of photos (the only one exception is Fast geometry reconstruction mode, which is limited to 128 photos due to algorithmic restrictions).

But the 3D reconstruction algorithms are very resource intensive, so the maximum count of photos is usually limited by the available RAM size and available computation time.

An effective limit on the count of photos is about 30-40 on a 32bit OS, and about 100 on the 64bit OS with 6Gb of RAM.

Here are the system parameters recommended for best performance, in the order of importance:

1. Number of cores (4 physical, 8 logical cores is recommended)
2. Amount of RAM (6 Gb is recommended)
3. Supported SSE version (SSE4.2 is recommended)

So for optimal performance I would recommend a Core i7 based system with 6Gb of RAM and 64bit OS.

We are planning to include an option to reconstruct the model in parts in the future versions of PhotoScan. Currently you can stitch partial reconstructions only in the external 3D editor.

Please also note, that PhotoScan is designed mainly for the general purpose 3D modeling, so we can't introduce professional GIS functionality in it, at least for the PhotoScan price level that is not so high.

With best regards,
Dmitry Semyonov
AgiSoft LLC
With best regards,
Dmitry Semyonov
Agisoft

Nathan Craig

  • Newbie
  • *
  • Posts: 18
    • View Profile
    • Personal Page
Re: Re: Image Collection Limits
« Reply #2 on: July 24, 2010, 07:48:18 PM »
Dear Dmitry,

Thank you for your detailed response. This will help me configure equipment that is most efficient for processing image collections with PhotoScaner. I am presently in the field, internet access is intermittent, and access to new hardware impossible. When I return to the US, I'll be looking into installing PhotoScaner on a high end Win64. I've got a pair of machines in my lab that would be perfect. In the mean time, I seem to be able to regularly process collections of >30 4 megapixel images. I'm considering downsampling a little more to see if that will permit more images to be included in the collections. In the mean time I have been working with the following process:

Select 25-30 images that are on a single "flight line" and process these images. Once the mesh and photo texture is created, I export this as a ply file. Then I process the next 25-30 images. I often leave an overlap of 2-5 images to ensure that both sections will have enough common form to permit co-registration.

When I open a pair of ply files created from different image sub-sets the scale in the two models looks "pretty close". In VRMesh Studio, I use a manual registration routine and then perform a global registration between the two sections. I've not tried registering two sections in MeshLab, but this is on my list of things to try.

The ability to process 100 images with 64 bit OS would be fantastic. That would significantly reduce the number of sections that have to be registered.

Do you have any suggestions to ensuring that the scale is similar in different sets of sub-collections? Given the way I place targets, I can't always be sure there are two separate known points for setting scale. If there are processing parameters that would help ensure common scale across sub-sets of a larger collection, I would be very interested to learn about this.

Thanks again, I've already been getting good use out of PhotoScan.

Dmitry Semyonov

  • Agisoft Technical Support
  • Full Member
  • *****
  • Posts: 200
    • View Profile
Re: Re: Re: Image Collection Limits
« Reply #3 on: July 25, 2010, 09:30:49 PM »
Hello, Nathan,

Unfortunately, there is no way to extract the scale from the photos alone. It is required to know the distance between any 2 points in the scene to set up the scale.

We are currently considering an option to perform automatic stitching of reconstructed fragments by feature points matching. That would provide the consistent scale between the fragments, but 2 known points for the whole set will be still required to set the absolute scale.

Currently, your way seems to be the only one.

I'll keep you informed about the development progress for the automatic photo sets alignment option.

With best regards,
Dmitry Semyonov
AgiSoft LLC
With best regards,
Dmitry Semyonov
Agisoft