Bundler Photogrammetry Package

By Josh Harle

This package is now very old. (things have moved really fast!)  I recommend using Changchang Wu’s VisualSFM, which works splendidly and continues to incorporate new research. 

 


Last week I held a workshop on Photogrammetry and Augmented Reality. As preparation for the workshop, I took the fantastic BundlerMatcher tool by Henri Astre, and packaged in a few tools to make it as simple as possible to use.

The Bundler Photogrammetry Package can be downloaded from http://files.neonascent.net/SFM.zip. A guide to using it is available at http://files.neonascent.net/photogrammetry-guide.pdf.

The package is a combination of a number of different programs and packages, along with some script of my own. The main advantage at the moment of this approach over Astre Henri’s other tool; PhotoSynth Toolkit, is that you can break image-sets into chunks to process in higher detail without crashing/running out of memory.

85 thoughts on “Bundler Photogrammetry Package”

  1. Wow, that’s great Josh. I was just about to do exactly the same thing since so many people have trouble setting it up. You beat me to it and saved me the work (phew).

  2. Pingback: PMVS2 x64 and videos tutorials » Visual-Experiments.com

    • @fc: Finding the CCD size can take some research for some cameras, but sometimes it’s as easy as typing ‘ “ccd size”‘ into google. The size is normally listen in the 1/3″ style, but it’s worth checking the extract_focal.pl file for other cameras that may be listed with that sensor format and copying down the mm value, rather than working it out yourself.

  4. Josh,

    Outstanding work as always. This is really impressive stuff. I started working through the tutorial, and I encountered an error while running the sparseRecon64.bat file. The process gets to the point where it states [Sift Feature matched] and then my machine returns an error. It says BundlerMatcher.exe has stopped working. Under the problem details I get the following:

    Problem Event Name: APCRASH
    Application Name: BundlerMatcher.exe
    Application Version: 0.0.0.0
    Application Timestamp: 4c8765a8
    Fault Module Name: atio6axx.dll
    Fault Module Version: 6.14.10.9921

    I click the dialogue to close the box and the batch file continues extracting exif tags. Bundler continues to run. When I look in the bundle folder there is an out file but there are no ply files. I’m running Win7 64 bit.

    I’m really looking forward to using these tools. Any suggestions?

  5. @Nathan:

    Regarding your error message, this error actually doesn’t effect the operation of the feature matcher – it occurs as the feature matcher is closing. It seems to be an overzealous windows error reporting system. You can disable it by following the instructions here: http://www.raymond.cc/blog/archives/2009/08/12/disable-program-has-stopped-working-error-dialog-in-windows-server-2008/ Specifically the last one involving running “regedit” and setting “HKEY_CURRENT_USERSoftware MicrosoftWindowsWindows ErrorReportingDontShowUI” to “1″.

    In remedying your lack of .ply files, I’d suggest (in order of effect):

    1) making sure the list.txt uses lowercase “jpg” not “JPG”, in rare cases MakeList decides to use JPG, which breaks the later script. Just run it again.

    2) adding the camera ccd size to the c:sfmbinextract_focal.pl file,

    3) modifying the sparseRecon64.bat configuration, setting
    BundlerMatcher list.txt gpu.matches.txt 0.8 0
    to
    BundlerMatcher list.txt gpu.matches.txt 0.8 -1
    if there are < 1000 feature points on average found currently, or
    BundlerMatcher list.txt gpu.matches.txt 0.8 1
    if there seem to be too many.

  6. @Josh

    Setting “HKEY_CURRENT_USERSoftware MicrosoftWindowsWindows ErrorReportingDontShowUI” to “1″ eliminated the error reporting.

    Thank you very much for the leads on getting ply files. At first, trying each of the three solutions listed did not produce positive results. This was very puzzling. Then I realized that I was working on a machine that had been converted from 32 to 64 bit. This conversion had caused some problems with other applications, and I got to wondering if it was not also the cause of my problems getting ply files to generate. I installed the toolkit on another machine that has only ever had a 64 bit OS. The toolkit ran like a charm. Problem completely resolved, the issue is with idiosyncrasies that arise from converting an old 32 bit box to 64.

    Josh, your toolkit is fantastic and your video tutorials are very clear. Thank you very much for sharing your knowledge with the community.

  7. @Nathan

    Very glad to hear it!

    Credit should go to Henri Astre who released the BundleMatcher tools, as well as Yasutaka Furukawa and Jean Ponce for PMVS2, Pierre Moulon for the Windows binary of CMVS and PMVS2, and Noah Snavely for Bundler; I only pulled some bits together and made scripts to avoid the command-line. :)

  8. Hi Josh,

    great job! But I have some problems because I am working on a 32 bit machine. Could you please tell me where I can find the file that you indicate in the tutorial for who uses a 32-bit system?
    Thanks a lot

  10. @pr and @fc:

    The 32-bit processing stuff that was in the package wasn’t any better/faster than Bundler, and was also not open-source. If you don’t have access to a 64-bit machine, you can use the Bundler software, though this is without the scripts to make things easier.

    Sorry for the inconvenience.

  11. Josh, what would I have to modify in order to run this from a drive letter other than C: ?

    I ask because the only 64bit PC that I have at my disposal serves its primary purpose as an iMac and the Boot Camp partition is just large enough to install Windows 7 and my core applications. My TEMP directory and virtual memory paging file are configured to external hard drives and all of my free disk space is also on external hard drives. I know that this decreases read|write speed, but I’ve run this way for about a year without incident.

    Currently my free space is all located on drives X and Y, so if there’s a simple way to modify a few strings from C to one of those, then I’d appreciate it. Needing to constantly copy an entire set of images to a separate processing folder, rather than just telling the script|application where the images’ current locations are, as Photosynth does, is somewhat of a pain point with me.

  12. @Nate

    There are two options that you could consider. The easiest is to process your image files from where ever you would like, but leave the base install at c:SFM. Assuming you have the ~60mb on the C: drive, you can just copy the scripts to any folder, and run them from there.

    Alternatively, You can put the SFM directory where ever you want, but will need to modify all of the scripts to point to the new location, and possibly some of the more hidden scripts that these point to. If you did this, it would probably be helpful to just search within files for “c:sfm” and replace.

    Obviously option #1 is the easiest, and this is what I do.

    Hope that helps!

  15. Thanks a lot for the package.
    I made some successful tests.
    I cannot understand very well DenseRecon settings about image segments, resolution and treshold.
    Another question is: can Bundler+Pmvs2 give pointclouds in real measures or do i need to scale them by aother softwares?
    Thanks

  16. I installed it at d:/SFM. From memory I only edited two files: sparseRecon64.bat and denseRecon_batch.bat. If you open each in notepad you’ll see the second line (path …) sets the search path for executables. Just change those drive letters. I’m pretty sure that was all.

  17. @Tim

    Yep that looks about right. I’ve just double-checked, as I had remembered there being absolute directory location in “2_RunBundler.gpu.sh” too, but it turns out not.

  20. By the way, Josh, last weekend after talking with you, I was running the toolkit and every time that BundlerMatcher got to 8% of feature extraction, it crashed. I’ve run out of memory with Photosynth many times during scene reconstruction, but never ever during feature extraction. Does BundlerMatcher try to keep all the features in RAM until it completes matching? If so, I’m not sure how I’m going to be able to get any larger resonctructions with Bundler than I get with Photosynth, even with the 64bit version of Bundler.

    Do you have any suggestions? I can completely understand if Bundler runs out of memory during scene reconstruction, but not feature extraction.

  21. @Carolina: That’s great to hear! and

    @Nate; I agree; it would be great to see what you’ve got, and have somewhere to share results.

    Feature Point extraction only deals with a single image at a time, so that’s quite odd to hear. How many feature points is it finding? I have had to troubleshoot problematic images before, so check which image number it is crashing at, and try removing that number and/or the next in the image sequence. It is probably something particular with that image.

  22. I’ll have to get back to you on the number of features at the time of the crash.

    So far, every time that I’ve launched BundlerMatcher with the toolkit, the first thing it does after I press spacebar to begin is print out this error message:

    [BundlerMatcher]
    [Initialization]Compile Log
    (1) : warning C7531: global type sampler2DRect requires “#extension GL_ARB_texture_rectangle : enable” before use
    (4) : warning C7011: implicit cast from “int” to “float”
    (5) : fatal error C9999: *** exception during compilation ***

    Does this have to do with my drivers not being up to date or my GPU not being up to snuff? (I’ve only got a GeForce 9400 in this thing and am not sure how much RAM it’s got onboard.)

    I’ve been trying several image sets tonight and all of them process most of the first hundred photos and then begin printing out error messages about not being able to read the next images through the end of the image sequence.

    One thing that has remained consistent throughout my usage of BundlerMatcher is the immediate consumption of all available RAM when it begins to run. I’ve never actually got to the feature matching because of all of my troubles getting feature extraction to complete.

    I’m currently re-running BundlerMatcher after removing the last image that it successfully processed, the photo preceding and the photo following, so I’ll see if that gets me any further, but this random rejection of images is not very pleasant. To my knowledge, they don’t differ in any way.

    • @ Nate

      I have had the compilation error before on certain setups, and despite it’s dire tone, it’s not actually fatal to the process. The rest of the symptoms though don’t sound great. Can you email me the output, so I can have a look at the errors? Or possibly an imageset that isn’t working?

  23. Hi, recently im having the problem of “no ply files output” even if i change the value inside sparseRecon64.bat. it seems to work after the third or fourth attempt with value -1. It’s very strange and it takes too long time (20 min for each attempt). BundlerMatcher process seems to work well (sometimes with the error “too many features..riallocate…”). Bundler process instead seems to be too quick (some seconds instead of some minutes) and i get no ply output. In order to which of two processes fails i would normaly run SparseRecon64 and If a don’t get ply output run only Bundler process again. How can i do that?
    Thanks.

  24. @marco,
    Download the newest SFM.zip bundle to repalce what you have. The new one has renamed sparseRecon64 to sparseRecon, so you’re using an older version. Also, if the first stage does not produce any .ply, then the dense recon won’t. To improve results, it’s worthwhile adding your camera details to the extract_focal.pl table (see the guide for details – or the video).

  26. @marco,
    Apologies, I had planned to rename for the latest release, but didn’t. Since you are using the current version, I’d say the camera parameters are the significant element.

  27. @josh
    ok. my camera parmeters are “NIKON CORPORATION NIKON D90″ => 23.6 I want to underline that after 3 or4 attempts i can get ply output.
    Sparserecon is made of two steps:1)bundlermatcher 2)Runbundler
    I realized that after all the “extracting exif from image”, the “RunBundler” process is too fast (2 or 3 seconds). Indeed it takes almost 30 sec when i get ply outputs. Maybe it’s a RunBundler process problem. i wondered, if i don’t get ply may i run only RunBundler process after deleting the fold “bundler”? how can i do that? When it works pmvs give me wonderful point cloud.

  28. @Josh
    Right now i’ve been solved changing the value -1 to 1 inside SparseRecon64. That’s very strange. Sometimes works with -1 and sometimes with 1. Four attempts. What’s the meaning of this value?
    I can’t undestand how to realize if it wil work during the processing itself. It seems to get the same output files without ply when it doesn’t work.
    Thanks.

  29. @Marco,
    Here are details of the parameters, taken from the c:sfmREADME_BundlerMatcher.txt:

    - last two parameters:
    0.8 -> matchThreshold : 0.0 means few match and 1.0 many match (float)
    -1 -> firstOctave: specify on which octave start sampling (int)
    firstOctave: low value (0) means many features and high value (2) means less feature

    An intermittent problem I’ve found is that sometimes the key files are saved as “.JPG.key” instead of “.jpg.key”. Check the /bundle/out file when it has failed, to see if it reports at the end that it can’t find the .jpg.key files. To remedy this, just make sure the first stage is creating “.jpg.key” files, or cancel and retry if it is not.

    • @Erik, It looks like the sparse recon did not produce anything ( no bundle.out file in the /bundle directory ) and so RadialUndistort crashes when it is pointed at files that don’t exist. Try running through the Kermit example images to make sure that it works with those.

  31. @Josh, I tried it with the ET example (was no Kermit example) and it worked like a charm (although there were no ccd information on the camera used). But when I tried it again with my test the same problem occurred.

  33. @Erik,
    If no images are matched, or there is a problem, the only output is the “out” file, with no “bundle.out”. These are no the same file. Open the “out” file in notepad and scroll to the end to see if there are any errors. As I mention above, sometimes the key files can be incorrectly generated as “.JPG.key” rather than “.jpg.key”.

    Also, it’s possible that your images are just not very easy to match. From the above troubleshooting advice:

    3) modifying the sparseRecon64.bat configuration, setting
    BundlerMatcher list.txt gpu.matches.txt 0.8 0
    to
    BundlerMatcher list.txt gpu.matches.txt 0.8 -1
    if there are < 1000 feature points on average found currently, or
    BundlerMatcher list.txt gpu.matches.txt 0.8 1
    if there seem to be too many.

  34. @Josh,
    Tried it again on another set, and it produced the files with .JPG instead of .jpg. But after a few more tries it worked, producing .jpg files. But after finishing that model and going back to the first one it only produces .JPG. Do you know why it does this? Is there something I can do to controll this, other than trying over and over again until it works?

  35. hi all, i got some very good models, up to 2 millions points. my problem now is how to “georeferencing” the pointcloud. I have 4 targets for each models, i acquired the coordinates by a total station. I tried to use “java graticule3d method” without success. The point cloud is to big for excel. could you suggest a software which let me to move the models to the right positions. The models are indipendent structures, i’d want to recreate all the set of structures. Thanks

  37. Marco and Josh, I’m not sure if this will handle the size of your dataset or not as I’ve not used it myself yet, but Jorn Anke recently posted something that I think is very much in line with what you’re interested in.

    http://diydrones.com/profiles/blogs/georeferencing-of-pointclouds

    There’s also some previous discussion here (as brakar):
    http://photosynth.net/discussion.aspx?cat=6b63cb81-8b57-4d5d-a978-41d5509bf59a&dis=a26a7660-1d0f-4f12-a39c-e6e952b7da99

    as well as here:
    http://photosynth.ning.com/forum/topics/gcp-in-the-point-cloud (which is where I got the first link from)

  39. I read about PC-Georef, it’s not redy for sale.
    I cannot think that there is any 3d software which can move and scale at the same time a model, pointcloud or mesh, basing on coordinates of some points. I saw different software as meshlab, vrmesh, geomagic, with align function between meshes or point clouds but no one is able to align a model basing on coordinates like in autocad. I hope i’m wrong.
    thanks

  40. Marco, I would just email or send a private message to Jorn, if I were you.

    From what I understood from his post, although he is not selling it yet, he is eager for people to test it for free to help make sure it works for more data formats.

  41. Hello,

    I have created the blender ‘ply’ files of ET using your software (32 bit, not 64 bit version)… the ones in the bundler directory are fine.

    When I run the DenseRecon.vbs using default values, the ply files in /pmvs/models has no points, but when I alter the values I get a dense mesh that looks like ET has exploded! It happens on other photo datasets I try to process too. The point cloud has colour, but blocks of points look nothing like ET or other things I process.

    Any ideas what may be wrong?

    Cheers, Tom

  42. Done.
    i’m trying to georeference meshes using Meshlab and autocad. I’m trying to extract from a mesh only the targets then i’ll georeference by autocad import again into meshlab. Then i’ll use the picking points tool (i’ve not understood well how to use it yet) to georeference the model. It’s just an idea (i’m a newby in 3d). i’ll let you know.
    Thanks

  43. Hi, im trying to use the photogrammetry kit for a project. I watched your video on it but got stuck on a really early stage. When I run “sparserecon64″ from inside the folder I created (4:09) the program seems to start but I quickly get an error message saying:

    ” ‘Bundlermatcher’ is not recognised as an internal or external command, operable program or batch file” Press any key to continue
    ‘sh’ is not recognised as an internal or external command, operable program or batch file” Press any key to continue ”

    I have no idea what this means. I’ve read a few different readme files in the download and they suggest I need to use cygwin (I can see the cygwin folder but I have no idea what it does or how to use it). If you have a spare moment can you give me some hints on getting sparserecon to work?

  45. Hi, back again with another hiccup. I’m having the problem mentioned above in which running “sparserecon_64″ doesnt seem to create any .ply files in the bundle folder. I do get a file called “out” that contains the following text:
    ———————————————————
    [OnInit] Running program C:sfmbinBundler.exe
    Loading images from file ‘list_focal.txt’
    [BundlerApp::OnInit] Processing options…
    [BundlerApp::OnInit] Loading frame…
    [BundlerApp::OnInit] Loading images…
    [LoadMatchTable] Loading matches
    [LoadMatchTable] Error opening file gpu.matches.txt for reading
    ———————————————————
    Does this last line make any sense? There is a file called “1_extract_GPU_features” in the SFM directory, but running it from there doesnt seem to do anything.

    Half way through running “sparserecon_64″ (just before the “hit any key to continue” message) I get a line saying:
    ———————————————————
    Error: cannot initialize openGL context for SiftGPU
    ———————————————————
    is this relevant?

    I’ve tried the 3 steps suggested to nathan on Oct 01. None of these seem to make a difference. The other posts about this issue seem to be specific to error messages that I’m not getting. I seem to be getting the correct EXIF settings. Any help would really be appreciated.

  46. @Sam;
    The error “Error: cannot initialize openGL context for SiftGPU” suggests you don’t have a compatible 3D graphics card. Does that seem likely? You can try Henri Astre’s SFMToolkit and see if that works. Otherwise you could look at using PhotoSynth to reconstruct, and pull the pointclouds out afterwards.

  47. Hi, I downloaded Agisoft Photoscan in the end. I can use that for most of my needs, including making a bundler.out file before using Denserecon to make a more detailed point cloud. Thanks for your help here. If a gallery gets sorted Ill try and upload some results.

  48. @Sam;
    I have also used PhotoScan similarly, though I’ve found the alignment time increases exponentially as your imageset size increases. For large sets it’s quite a time investment.

  51. This looks great so far, but am having some trouble setting the CCD size for my camera. I added my camera name and ccd size to the file extract_focal.pl, but the program still says that no ccd size was found and no .ply files are generated. Am I missing something simple on how to set it? Thanks

    • @Dan, a few things to check: 1) That your images have Make and Model details in the EXIF data (i.e. look at the file Properties, Details in explorer), and 2) that the camera name in extract_focal.pl is exactly the same, including capitalization.

      You can get an indication of the formatting by looking at the other camera names in the file.

      Hope that sorts it!

  52. is this mac compatible at all? using wine bottles or ?? i would really be interested in using this if that was possible [without boot camp]

    • @Jon,
      I’ve used Photoscan quite a bit, and it certainly has its place for reconstruction. It will match and position all images in a set, which is sometimes an advantage, and sometimes problem if there are some outlier images that don’t work well.

  54. Hi Josh, Thanks for putting this together into a nice package. I played around with it back in December and got some great results. I ran a huge dataset (784 images from a 12MB camera) on a 64-bit machine with 48GB RAM and an nVidia Quadro FX 5800 (4GB) graphics card. It took it a while to process, but I let it run over the weekend and logged in remotely to hit enter from time to time :)

    With regards to Sam’s issue from Dec 21, I was getting the same error today when trying to process another dataset (smaller!) on the above machine but logged in through remote desktop. When I went to the physical machine and ran spareseRecon64.bat file the error disappeared. Not sure if Sam was running remotely, but I thought I’d point this out in case others experience the same problem.

    This is the error I was getting when I ran spareseRecon64.bat:
    [BundlerMatcher]
    [Initialization]Shader not supported by your hardware!
    Error : can not initialize opengl context for SiftGPU

  55. Hi Josh i use the ET file in the example folder and it also just produce file “out” without any extension after i run the sparserecon64.bat I’m not sure why it happened. Am I missing something?
    second, is it possible to use any picture from internet like google to build model?as we know there are so many different camera in google.
    thanks

    • @Kazam,
      Have a look at some of the trouble-shooting conversations already in the comments. I’m not sure what the likely problem is at the moment, especially if there are no error messages. I can guarantee that the ET images reconstruct okay.

      You might want to look at using Henri Astre’s SFMToolkit, which is being very actively maintained and improved.

  56. I’ve same issue as Kazam. THe sparse recon fails, says it can a memory adress (large number) and then the process seems to continues but bundler directory is empty.

    I run an Ati HD 4850 wich I believe is the problem.

    I really want to try to reconstruct large scenes but I keep failing.. Tested the Henri Astres toolkit too…

  57. Hi,

    Great tutorial, worked first time! I was able to get the ET tutorial to work. Then I made a folder for my first set of images and got the following error:

    [BundlerMatcher]
    [Initialization] Too many features, reallocate texture, Too many features, reallocate texture

    Any ideas as to what this problem is?

    Thanks,

    S

    • @Scott,

      Edit the sparseRecon64.bat and change the value
      BundlerMatcher list.txt gpu.matches.txt 0.8 -1
      to
      BundlerMatcher list.txt gpu.matches.txt 0.8 0

      this will reduce the number of feature-points searched for.

      Hope that helps!

  59. @Scott,
    try
    BundlerMatcher list.txt gpu.matches.txt 0.8 1

    Depending on how big your source images are, you may try reducing their sizes too if that doesn’t work.

    The good news is that this probably means they have lots of good points for matching.

  60. Pingback: Facets | neonascent

  61. Hi, greatings. I am running the ET example in a 12GB Ram, windows7 64bit, with a Quadro 3800 and 8 cores (2x4cores xeon). All worked ok, but the final step of denserecon is still working after 5 hours?? I just checked that the photo directory is in C:SFMET (not c:SFMexamples|ET since i made a copy there so i preserved the original directory). How much time should the ET example take?
    Victor

    • @Victor, 5 hours is probably too long even for the level 0, csize 1 reconstruction. Try with level 2, csize 2 and it should just be a few minutes (less than 5).

  62. Hi. Good News. It seems it had some kind of bad neighborhood with another software that we had running at the same time. It worked ok already, and since then we have been processing several different tests (a chair, a stone sculpture and a column – around 10 to 30 photos each). We are playing with the threshold parameters a bit to see what happens (0.5 is better quality than 0.9 or worse quality? 0.5 seems to give more points in the end). Strangely, sometimes the sparse works but in the end it doesn’t output the ply files… but we then run it a couple of times more and suddenly, without we changing nothing, it works ok and outputs the model files! We will try to put the results online so you can see the output. Thanks for making the package available!

  64. Hi again! I have here a strange error. The process runs ok despite this error, but i am wondering if i am skeeping some important step :-( .
    The error appears when we run “desnserecon”:
    “[WriteBundleFile] writing 14 images and 2846 points…
    undistorting image IMG_3037.jpg
    (it does this ok for all the images then the error appears in the next line)
    /cygdrive/c/sfm/bin/RadialUndistort: /cygdrive/c/sfm/bin/RadialUndistort: cannot execute binary file
    Reading bundle… 14 cameras — 2846 points in bundle file”
    And gos on until it processes everything
    I checked and the files are there: both “RadialUndistort” without extension and “Radialundistort.exe”. Could be a problem with the file or with the path?

  66. Hey Josh,

    I’ve installed the bundler package, but when I run the example, I only get the out file int he bundler folder. And when I tried it on my older computer, it would give me;
    bundler.init.out
    bundle.out
    bundle_001.out
    out
    points001.ply

    I have 6 images, about 2600 x 3908 of a hand. They are jpgs converted from CR2s. I’ve read through every comment on this page, and can’t find anything. Please help!!! :)

    Thanks!

    J

    • @Jay,

      Have you tried running the process on the sample ET imageset? This will give an indication if anything is going awry in the setup, or if it’s the images.

      For a good indication of whether your images will reconstruct well, run them through PhotoSynth (I have a tutorial video on vimeo).

      You may need to modify the sparseRecon64.bat to change the reconstruction configuration (with such large images you might have to use “BundlerMatcher list.txt gpu.matches.txt 0.8 1″

  67. Hey Josh,

    I ran the ET images with -1 and only got;
    bundler.init.out
    bundle.out
    bundle_001.out
    out
    points001.ply

    And when I changed it from 0.8 1, I would only get the “out” file.

    Any ideas?

    Thanks Josh!

    J

  68. ran the ET images with -1 and only got;
    bundler.init.out
    bundle.out
    bundle_001.out
    out
    points001.ply

    And when I changed it from 0.8 1, I would only get the “out” file.

    Any ideas?

    Thanks Josh!

    i also meet the same question as Jay

    can you give me some advice ?

  70. This package is now very old. (things have moved really fast!) I recommend using Changchang Wu’s VisualSFM tool, which works splendidly and continues to incorporate new research into the tool.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>