Wednesday 17 October 2012

How many pictures could my laptop elaborate with the SfM/IBM technique (using PPT)?

Hi all,
I'm testing the potentialities of the SfM/IBM technique, progressively increasing the number of the pictures. It is my personal Guinness World Records. In the beginning my record was 100 images, then 105 and now ... 120.

I'm using the following hardware:

LAPTOP
Acer Aspire 5950G
Intel Core i7-2630QM 2GHz
8GB RAM
AMD Radeon HD 6850M

DIGITAL CAMERA
Nikon D70
3008 x 2000px

To reach the new record my laptop needed more or less 60 hours of calculation (maximum image resolution), but my work was only five minutes, the time i need to run Bundler and PMVS2 in PPT.

The result are amazing :)

The picture below show the structure i try to rebuild and ...


... the pictures below are some screenshot of the final model.




Have Fun!

12 comments:

  1. Hello,
    Nice job ! The main issue with (PMVS) patch based reconstruction is that all is loaded in RAM. So memory is always the bottle neck, not CPU. For exemple i think 8go is not enough to manage 100 pictures at full resolution, so you may have swapped (it can explain the 60 hours of calculation). For quasi planar object like this, (it's basically 4 wall, so 4 quasi planar objects) I should advice you to Use MicMac, which is based on depth map calculation, from the French Geographical Institut http://www.micmac.ign.fr/.

    ReplyDelete
    Replies
    1. Hi Romain,
      thanks for the advice. We alredy tested MicMac (http://arc-team-open-research.blogspot.it/2012/03/example-test-of-micmac.html), but just with the default data. We were waiting for a motivation to test further this software and your comment is a good reason to try MicMac with quasi planar objects and see the difference.

      Delete
    2. Hi Romain,
      i hope to have time to restart to test MicMac during next winter. i could try to use the same photoset and to compare the time consuming and the final result. if i'll be able, i'll write a new post on comparison.
      thanks for your advice

      Delete
    3. Hi,
      i control the "vis.dat" and i discovered that PMVS2 did not use 36 photos. In the end Bundler worked with 120 images and PMVS2 with 84.

      Delete
    4. Yes, in the pipeline, it's the job of cmvs to choose best images. Then pmvs do the match.
      Maybe if you send me your photo set i can try to make do something with MicMac. I could also help you to go deeper with this wondeful software if you need.

      cheers,
      Romain

      Delete
    5. Hi,
      the images i used are here: www.museidironzone.it/scambio_file/spw121012027_sfm_ext.ply.tar.gz.
      Yes, please. If you can help me to go deeper in the knowledge of MicMac would be nice! Thank you very much, Alessandro

      Delete
  2. Amazing results! I never passed the 70 images limit (1200 px). But sometimes is better to create a single model using point cloud results from different PPT processes. You can waste lots of time processing a large image set with RunBundler and finally discover that RunCMVS can only use a small number of those images (i.e. blurry photos, not enough common features). I'm not pretty sure, but apparently when feature matching phase is running, things also start to get incrementally slow. More images = more slow.

    rgaidao

    ReplyDelete
    Replies
    1. Hi Ricardo,
      you're right, but in our case is often better to waste the computer time (leaving the computer working alone) rather than wasting our time (georeferencing many pointclouds). One of the thing I like about SfM and IBM is the few interaction the software need, so while the computer is processing the data, you can work on other things. But this is just my opinion :)

      Delete
    2. Hi Ricardo,
      thank. i have the same opinion in "incrementally slow. More images = more slow". if i remember well, with 100 images the time of calculation of PPT was only 48 hours.
      i will make a new test elaborating the inner floor of the same building (other 120 photos) and control again the time. Maybe there was a problem with the first photoset.

      Delete
  3. FYI i've made a rudimentary (for now, but i don't have many time to continue to work on it) WEBGL based point cloud viewer vizu.herokuapp.com. It's build upon the play framework (server side) and three.js (client side). You need either chrome or firefox and a recent graphic card to enjoy the full (very rough !) experience. right click, left click, ctrl and alt are your friend.

    ReplyDelete
    Replies
    1. Hi Romain,
      it is a very nice work. we were looking for a software like your to integrate in ArcheOS. Do you know the license you will use to release your software? we are trying OpenWebGlobe (http://swiss3d.openwebglobe.org/).
      thanks

      Delete
  4. I think it will be release under GPL2 or CeCill. I'm going to write you an email for further information.
    Thanks,
    Romain

    ReplyDelete

BlogItalia - La directory italiana dei blog Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.