Thanks for this tutorial. There's so many unanswered questions, but I realise that this tutorial would have been LONG if you attempted to dive in any deeper on any of the topics. You provided just the base to get us through, keeping this tutorial as short as possible. Just what I needed.
I did! BTW... if you ever feel inspired to do a more indepth video, that gets in to the "why"s, I'd be very interested to see that. With so many settings, I'm quite curious to learn more about how to tweak them in order to get different results... particularly with COLMAP.
I was just waiting for that crash haha. I had NO idea you could do this with free software, times sure have changed, thanks for introducing me to COLMAP and Krita! Quick questions, does meshlab have any form of undo yet? And I noticed you're running a ramdisk as the working folder - have you noticed significant gains over a decent SSD or was it 'just because'? Thanks, really appreciate the tutorial!
Really glad you liked it :) Actually I don't think COLMAP has an undo system, its kind of a one way process. RamDisk are faster then SSD but I just like having some temporary space I can mess about with.
Fantastic this is one of the few examples of tutorial for 3dscanning using open source or free tools.. I have been having a getting some pretty good meshes and dense point clouds using open mvs after colmap.
I can't seem to get Colmap to really work, I have 64 images and it gets through 41, makes 1400 points and then just throws a error "No good initial image pair found." Even though I have multiple photo's of the object, from different sides and distances. And the 1400 points it did make, don't resemble my images at all.
That sucks :( the most common reasons for that are: * Viewpoint changes are too large * Not enough images * Blurry images * Lack of "texture" in the scene * Too many reflective surfaces
There is a troube when I using a dataset including 150 images to reconstruct the model, is there any other way to deal with the mask all? I've tried some wibesites to do this, but the AI remove bg always goes wrong. forgive my poor ENG. Looking forward to your reply.
new question: I spent a lot of time to remove the background of these images, but when I use colmap to densify Construction, the result of construction still have background, which means it was a wasted effort.
Has somebody had a problem like, I would like to create dense > stereo, and a program quits after a few itteration... (I set the tdr value in nvdia Nsight ti 120seconds) (xeon, 16Gb Ram, Quadro P2000 HP workstatin Z 440)
I thought the algorithms use background as a reference to place the object in 3d model space. Please could somebody more familiar with the subject confirm?
Blender also has a decimate modifier to reduce the number of polygons. It is certainly not as good as doing it by hand but it should certainly tried before doing anything by hand. Who knows, you might get lucky and get a rather good low poly model.
Great video, also is great that you're only using open source software, but with all the time you spent on masking those pics you can actually sculpt a rock from scrath I think.
Maybe you can create an action in your image editor that extracts the background, like a droplet batch in PS or something. Agis photoscan is pretty fast when it comes to mask the images but I guess the point of this is staying with open source.
Do you remember what type of computer you did this on? I haven't had success with the other Meshroom app, because it crashes at the point of meshing it, so I am trying Colmap now. I'm trying to do it on a t460p Thinkpad Lenovo laptop, any suggestions on how to do a simple and less heavy model?
if i used the stereo,and the software instantly backing then quit, can you tell me what's wrong with my colmap software??? how to solve this problem? waiting heartily for your reply
Whenever I try to apply "parameterization + texturing from registered rasters" Meshlab just stalls and stops responding. However, the only difference I have made to my project is when I apply my settings for "Screened Poisson Surface Reconstruction". I set my reconstruction depth to a 15 instead of 9 or else it looks like garbage. I do this because it's a mapping of a large outside area. Several acres with buildings/structures and trees. Any ideas how I can get around this?
@@ThePassiveAggressor I think I also might try deleting the faces that are created after I convert it into a mesh. The terrain creates a large wavey plane, even after I've already deleted the excessive vertices. Don't know why I didn't think of that before. That could be a big help.
Great tutorial ! I have one problem tho. When i import my bundler to meshlab and apply textures to the mesh they are misaligned and weird looking. Is there any way to fix it ?
@@ThePassiveAggressor I managed to fix that problem by exporting Bundler from /dense/0/sparse using model_converter in command-line Interface. Thank you for your fast response tho !
Hi, I'm having trouble around 13:05 whereby when I try to parameterize + texture it says 'Failure: You need to have at least one valid raster layer in your project' but it wont let me rasterize the layer. Could you please help?
I'm not expert on MeshLab but make sure when you load the bundle into meshlab you get all your camera image in the right panel (they come in as raster layers). If not check your file structure and the bundle export. hope that helps
@@ThePassiveAggressor That worked perfectly thank you, the issue now is that I'm getting a black texture when I paramaterize and texture. Do you know any fix to this?
@@ThePassiveAggressor Yes, this was all done but it doesn't seem to want to work for me. I'm unsure of why this is but I've noticed that in the wire-frame mode you can visualize your textures from previous but it wont apply them onto the object itself. I'm unsure of what I've done wrong as I've followed your tutorial exact around 10 times to do the same thing but no results are yet to be seen. I can happily screenshot it all for you if you'd like?
Great tutorial, but I don't know why, when I try to perform parametrization + texturing from registered rasters, I get black texture ( I have ca. 80 rasters in set )
Fantastic video which finally covers al areas.Question, I have bee using Agisoft Photoscan to build assets and my aim is to map and import the entire large asset into unity as a VR cloud open world. however... how would you suggest I map such a large area of say 1km litrolerly square of a famous location here in Singapore?I thought of combining photos, video, using a fab pro 2's Kinect with Tango software. The issue is with all of these is just that, all of these... Agisoft wont like more than 1 focal length, nore does it accept lidar style/ flight of capture data, so its looking more likely that I will need to switch over to reality capture which accepts all without issue (kinda answering m own question) issue with this is needing to learn yet another soft/ pay for another software. I like the approach your taking but its very time consuming and for a little asset too... The nice thing about paid software is that is is designed to auto perform the tasks you have had to perform manually but it still looks pretty cool. (I will still lean this method).My end goal is 1 to 1 photo realism, the closest thing I come to in achieving this is 360 image capture and using virtual tour software which is more google maps/street view that The Matrix meets Ready player One which Im really going for.Any suggestions/ advice please. Sorry for the long message and thanks for any feedback you may give, feel free to pm me at Armstrong.skyview@gmail.comRegardsChris A
That's a great question! I too would like to scan large areas (to re-visit places I've seen on holiday). I know of a new piece of free software called Meshroom that automates more of the process, although I haven't had time to try it. Find it here: alicevision.github.io/ and let me know how you get on. Cheers
You mean in blender yes? I think there are several reason that can happen in blender, all I can suggest is try googling white texture in blender. sorry I couldn't help more.
I'm starting with Photogrammetie and i testet Visual FSM. But now i will try COLMAP :) Nice job and helpful :) So i'm using Blender 2.79 and for testing 2.80 ;) But i think 2.8 will crash :D
In my experience, and that of everyone on my team, Blender Internal has proven to be much faster for baking across texture (color) and normal maps. Give it a test yourself and let us know. Also, from months of testing, I can state that the accuracy of the baking of these maps is nearly perfect, even with the default settings.
Thanks for the tutorial! What are your PC specs? I used Meshlab for the first time the other day on my Macbook Pro from late 2014 and it was incredibly slow despite its 16GB RAM. Shitty Intel Iris integrated graphics card of course…
Yes, that is definitely an option but you'll always get better results doing it by hand. There also a great free bit of software called Instant Meshes that can do it but by hand is always better.
It's actual quite easy, go to github.com/colmap/colmap/releases press on the latest version and download the appropriate zip file, extract that somewhere then click on COLMAP.bat
Your welcome :) Easier yes but the nice things about taking screenshot by hand is that you pick the frames that will give the best result.(and only generate the files on disk for just those images)
When im using parameterization + texturing from registered rasters, the error "Failure of filter: 'Parameterization + texturing from registered rasters' You need to have at least one valid raster layer in your project, to apply this filter" shows up. Anyone knows how to fix this or what Im doing wrong?
Open your out.list.txt file in Notepad and see if it is empty. If this is the case you have exported the bundler files from the dense model. Load the sparse model again and save the bundler files from there and it should work. In Meshlab your images should appear in the Layer Dialog window after you open the project. Hope this helps
Theres quite the difference, Photogrammetry makes textured polygon models (usable in a lot of software) where as Gaussian splats make volumetric data that currently has limited uses (but is often more visually pleasing)
An excellent tutorial but my pictures weren't up to it, possibly because of my macro lens. Luckily tomorrow we get a fresh load of light so can try it again.
So I had the same crash as you, and managed to get enough debug data to make a good report on it. It's been found and will be fixed - though who knows when the next release will be, for people who don't compile it themselves. github.com/cnr-isti-vclab/meshlab/issues/500
I vote for no music! The video is really helpful - thanks! But the music isn't just unnecessary, it's a real distraction and at times swells to challenge hearing clearly what you're saying. Whatever music you put on, what type of percentage of viewers would like it and want it there? Very few I expect. And if I want some background music while following along, I've plenty of my own to choose from - and it will keep playing nicely regardless of going back over bits, pausing, and changing speed, etc!
Finally a tutorial that shows the entire process from start to finish, thank you so much!
Your very welcome :) glad it helped.
Best tutorial on photogrammetry and retopology on youtube
Thank you for the kind comment, its great to know my videos are helpful :)
Yeah
Best colmap tutorial! Will recommend!
Thanks very much :)
Thanks for this tutorial. There's so many unanswered questions, but I realise that this tutorial would have been LONG if you attempted to dive in any deeper on any of the topics. You provided just the base to get us through, keeping this tutorial as short as possible. Just what I needed.
Your very welcome, I'm glad you found it useful :)
I did! BTW... if you ever feel inspired to do a more indepth video, that gets in to the "why"s, I'd be very interested to see that. With so many settings, I'm quite curious to learn more about how to tweak them in order to get different results... particularly with COLMAP.
I was just waiting for that crash haha. I had NO idea you could do this with free software, times sure have changed, thanks for introducing me to COLMAP and Krita! Quick questions, does meshlab have any form of undo yet? And I noticed you're running a ramdisk as the working folder - have you noticed significant gains over a decent SSD or was it 'just because'? Thanks, really appreciate the tutorial!
Really glad you liked it :) Actually I don't think COLMAP has an undo system, its kind of a one way process. RamDisk are faster then SSD but I just like having some temporary space I can mess about with.
Fantastic this is one of the few examples of tutorial for 3dscanning using open source or free tools.. I have been having a getting some pretty good meshes and dense point clouds using open mvs after colmap.
Thanks :) I've heard other say open mvs is good too.
Amazing video mate, good voice quality, relaxing music and really well explained all with some tips.
Thank you for the nice comment, glad you liked it :)
this was brilliant! Also, that music was so relaxing. I kinda get a Final Fantasy vibe.
Very glad you liked it :)
It calmed me down every time Meshlab crashed or something broke.
Hey! When I click distortion in dense reconstruction, no images show up. Would u know a possible fix?
Not sure, try the official FAQ here colmap.github.io/faq.html
Cracking video. Absolute gem, thanks!
Thanks very much, I'm really pleased you've found it helpful :)
The tutorial really helps me a lot.
Very glad to hear it :)
Incredible!
Thanks :)
I can't seem to get Colmap to really work, I have 64 images and it gets through 41, makes 1400 points and then just throws a error "No good initial image pair found."
Even though I have multiple photo's of the object, from different sides and distances.
And the 1400 points it did make, don't resemble my images at all.
That sucks :( the most common reasons for that are:
* Viewpoint changes are too large
* Not enough images
* Blurry images
* Lack of "texture" in the scene
* Too many reflective surfaces
There is a troube when I using a dataset including 150 images to reconstruct the model, is there any other way to deal with the mask all? I've tried some wibesites to do this, but the AI remove bg always goes wrong.
forgive my poor ENG. Looking forward to your reply.
new question: I spent a lot of time to remove the background of these images, but when I use colmap to densify Construction, the result of construction still have background, which means it was a wasted effort.
@@caeoonshen5447 Can you use video software like Davinci Resolve/AfterFX/Apple Motion to "Rotoscope" out the background
Has somebody had a problem like, I would like to create dense > stereo, and a program quits after a few itteration... (I set the tdr value in nvdia Nsight ti 120seconds) (xeon, 16Gb Ram, Quadro P2000 HP workstatin Z 440)
I thought the algorithms use background as a reference to place the object in 3d model space. Please could somebody more familiar with the subject confirm?
Blender also has a decimate modifier to reduce the number of polygons. It is certainly not as good as doing it by hand but it should certainly tried before doing anything by hand. Who knows, you might get lucky and get a rather good low poly model.
That's a good point, thanks :)
Great video, also is great that you're only using open source software, but with all the time you spent on masking those pics you can actually sculpt a rock from scrath I think.
Thank's Miki, glad you liked it. It didn't take too long but there may be other ways to automate that like Rotoscoping
Maybe you can create an action in your image editor that extracts the background, like a droplet batch in PS or something. Agis photoscan is pretty fast when it comes to mask the images but I guess the point of this is staying with open source.
That's an interesting idea, thanks.
incredible tutorial !!! Many thanks !!!
Your very welcome, glad you liked it :)
So helpful, thanks for making this.
Your very welcome, glad it helped
Do you remember what type of computer you did this on? I haven't had success with the other Meshroom app, because it crashes at the point of meshing it, so I am trying Colmap now. I'm trying to do it on a t460p Thinkpad Lenovo laptop, any suggestions on how to do a simple and less heavy model?
It was quite a good desktop i5 I think. you probably need to download the version without CUDA so it will be quite slow
If I use parameterization + texturing from registered rasters, it results in all black textures. Any idea why??
I'm not sure but be sure you put yours images in a folder structure like I did.
if i used the stereo,and the software instantly backing then quit, can you tell me what's wrong with my colmap software??? how to solve this problem? waiting heartily for your reply
Not sure, check you have a good enough graphics card and the latest drivers.
Whenever I try to apply "parameterization + texturing from registered rasters" Meshlab just stalls and stops responding. However, the only difference I have made to my project is when I apply my settings for "Screened Poisson Surface Reconstruction". I set my reconstruction depth to a 15 instead of 9 or else it looks like garbage. I do this because it's a mapping of a large outside area. Several acres with buildings/structures and trees. Any ideas how I can get around this?
Sorry I haven't tried mapping a large area (although I'd like to try). Perhaps you can split it into smaller areas?
@@ThePassiveAggressor That's not a bad idea. I'll have to try that. Thanks!
@@ThePassiveAggressor I think I also might try deleting the faces that are created after I convert it into a mesh. The terrain creates a large wavey plane, even after I've already deleted the excessive vertices. Don't know why I didn't think of that before. That could be a big help.
@@BootyFebacca Great! good luck :)
Great tutorial ! I have one problem tho. When i import my bundler to meshlab and apply textures to the mesh they are misaligned and weird looking. Is there any way to fix it ?
Hum, I haven't seen that problem. May be worth checking the colmap / meshlab forums on that one. Sorry I couldn't be more help
@@ThePassiveAggressor I managed to fix that problem by exporting Bundler from /dense/0/sparse using model_converter in command-line Interface. Thank you for your fast response tho !
@@MrKajko112 Great! thanks for putting that info here for other to see :)
Neither of my machines supports CUDA, so I can't use Colmap for the dense point cloud before importing it to Meshlab. Any suggestions?
You could try VisualSFM, I havent tried it but there is a version without CUDA
Thanks for the suggestion! Working on it...
Yeah! Got dense points! No onto Meshlab. Thank you!
Thats great news, glad it worked :)
Hi,
I'm having trouble around 13:05 whereby when I try to parameterize + texture it says 'Failure: You need to have at least one valid raster layer in your project' but it wont let me rasterize the layer. Could you please help?
I'm not expert on MeshLab but make sure when you load the bundle into meshlab you get all your camera image in the right panel (they come in as raster layers). If not check your file structure and the bundle export. hope that helps
@@ThePassiveAggressor That worked perfectly thank you, the issue now is that I'm getting a black texture when I paramaterize and texture. Do you know any fix to this?
@@Milkeeyy humm not sure, make sure you set the vertex colours to white before hand, and check that meshlab can "see" your textures correctly
@@ThePassiveAggressor Yes, this was all done but it doesn't seem to want to work for me. I'm unsure of why this is but I've noticed that in the wire-frame mode you can visualize your textures from previous but it wont apply them onto the object itself. I'm unsure of what I've done wrong as I've followed your tutorial exact around 10 times to do the same thing but no results are yet to be seen. I can happily screenshot it all for you if you'd like?
Amazing
Thanks, glad you like it :)
Super ! thank you so much mister !
Your most welcome :)
Great tutorial, but I don't know why, when I try to perform parametrization + texturing from registered rasters, I get black texture ( I have ca. 80 rasters in set )
I not sure but I did have problems once when I didn't follow that exact folder / file structure, try checking that.
Fantastic video which finally covers al areas.Question, I have bee using Agisoft Photoscan to build assets and my aim is to map and import the entire large asset into unity as a VR cloud open world. however... how would you suggest I map such a large area of say 1km litrolerly square of a famous location here in Singapore?I thought of combining photos, video, using a fab pro 2's Kinect with Tango software. The issue is with all of these is just that, all of these... Agisoft wont like more than 1 focal length, nore does it accept lidar style/ flight of capture data, so its looking more likely that I will need to switch over to reality capture which accepts all without issue (kinda answering m own question) issue with this is needing to learn yet another soft/ pay for another software. I like the approach your taking but its very time consuming and for a little asset too... The nice thing about paid software is that is is designed to auto perform the tasks you have had to perform manually but it still looks pretty cool. (I will still lean this method).My end goal is 1 to 1 photo realism, the closest thing I come to in achieving this is 360 image capture and using virtual tour software which is more google maps/street view that The Matrix meets Ready player One which Im really going for.Any suggestions/ advice please. Sorry for the long message and thanks for any feedback you may give, feel free to pm me at Armstrong.skyview@gmail.comRegardsChris A
That's a great question! I too would like to scan large areas (to re-visit places I've seen on holiday). I know of a new piece of free software called Meshroom that automates more of the process, although I haven't had time to try it. Find it here: alicevision.github.io/ and let me know how you get on. Cheers
hey, I exactly followed the steps but if I go to texture view nothing changes. It's still all white! Any ideas what to do in this case?
You mean in blender yes? I think there are several reason that can happen in blender, all I can suggest is try googling white texture in blender. sorry I couldn't help more.
yeah in blender. alright I will do
Thank you. It seems creating that low res rock is completely manual and very tedious, I wonder if there are automated tools available.
Yes there's a program for that called instant meshes github.com/wjakob/instant-meshes
in my mashlab colors are not warking only white model is showing
Make sure when you load your bundle file you see a list of images on the right panel (like ruclips.net/video/bDHJM6nAKtc/видео.html)
I just can't finish.. I'm stuck with a CUDA problem in the dense reonstruction/ stereo option :S :(((
Sorry to hear that, you could always try the colmap support: colmap.github.io/
I'm starting with Photogrammetie and i testet Visual FSM. But now i will try COLMAP :) Nice job and helpful :) So i'm using Blender 2.79 and for testing 2.80 ;) But i think 2.8 will crash :D
Thanks, glad you found it helpful. I can't wait for blender 2.8 but I'm waiting for the proper release
Any special reason to use Blender Internal instead of Cycles for baking?
In my experience, and that of everyone on my team, Blender Internal has proven to be much faster for baking across texture (color) and normal maps. Give it a test yourself and let us know. Also, from months of testing, I can state that the accuracy of the baking of these maps is nearly perfect, even with the default settings.
Thanks for the tutorial! What are your PC specs? I used Meshlab for the first time the other day on my Macbook Pro from late 2014 and it was incredibly slow despite its 16GB RAM. Shitty Intel Iris integrated graphics card of course…
I have a pretty good PC, i5 with geforce 1080. COLMAP works best with a GPU with CUDA cores so sadly it will be very slow on Intel Iris :-/
I just quit using openmvs and meshlab to reconstruct large scale scene......
thank you
your very welcome :)
To make a low resolution of the mesh, you can reduce the number of polygonf with decimate tools on meshlab isn´t?
Yes, that is definitely an option but you'll always get better results doing it by hand. There also a great free bit of software called Instant Meshes that can do it but by hand is always better.
there are the same modifier in Blender software, called "decimate". Maybe ThePA are right about hand work.
And, thanks to author, very good tutorial!
very nice!! It would be great if you do a 'how to install colmap' tutorial for windows 10.
It's actual quite easy, go to github.com/colmap/colmap/releases press on the latest version and download the appropriate zip file, extract that somewhere then click on COLMAP.bat
hey! thanks for this tutorial! in't it easier to just turn video in to img sequence and work from there instead of doing screenshots?
Your welcome :)
Easier yes but the nice things about taking screenshot by hand is that you pick the frames that will give the best result.(and only generate the files on disk for just those images)
at the stereo stage after a few minutes my colmap turns off. Any advice?
oh dear :-/ if its crashing I guess check you have a good enough graphics card and enough RAM
@@ThePassiveAggressor I have 16gb RAM and a 2gb graphics card. I mount high definition movies and photos without a problem on my computer
@@patricko987 That sounds plenty good enough, try checking colmap.github.io/faq.html
Thanks a lot fot this!
Your very welcome! :)
Amazing!
Thanks :)
Couldn't you just reduce polygon number on the rock in Blender and leave the old texture at the same time?
Yes, you can get blender to reduce the polygons automatically but it will never be as good as doing it by hand. (but it does an ok job)
Fantastic. What was the reason for importing the Bundler data into Meshlab?
It brings in all the required camera data
EXCELLENT
Thanks :)
When im using parameterization + texturing from registered rasters, the error
"Failure of filter: 'Parameterization + texturing from registered rasters'
You need to have at least one valid raster layer in your project, to apply this filter" shows up. Anyone knows how to fix this or what Im doing wrong?
Bump, same error
Same
Open your out.list.txt file in Notepad and see if it is empty. If this is the case you have exported the bundler files from the dense model. Load the sparse model again and save the bundler files from there and it should work. In Meshlab your images should appear in the Layer Dialog window after you open the project. Hope this helps
I'm also having this same error. Anyone have a definitive fix?
Thank's man :)
your welcome :)
A free composting program like Fusion or Natron would make your masking go by MUCH quicker.
That's a very good point, I will have to take a look at that.
It is amazing.
Thank you :)
What diff between this and gaussian splat.
Theres quite the difference, Photogrammetry makes textured polygon models (usable in a lot of software) where as Gaussian splats make volumetric data that currently has limited uses (but is often more visually pleasing)
An excellent tutorial but my pictures weren't up to it, possibly because of my macro lens. Luckily tomorrow we get a fresh load of light so can try it again.
Thanks :) yes its important to have crisps images.
Man when I tried to get a model, it just used the ground from my images and there is nothing from the subject I was photographing
oh bad luck.. you might have needed more images, did you mask out the images?
Is a gpu required for the Dense reconstruction? I
With this setup I believe so
Thanks.
If you have issues with Stereo crashing your program like me. Follow this video. And restart your computer.
ruclips.net/video/iRel8fQRDz0/видео.html
Thanks for the info chase :)
your a god
Your very kind, glad it helped :)
i like this tutorial have u try use a Convert Video Into An Image Sequence
Thanks :) . I haven't but if you convert the video to an image sequence you will have 100's if not 1000's of images to look through.
What are your computer specs?
3.6ghz i7 with 12gb ram and geforce 780ti
Cool tutorial, but what the heck is up with the background music? We are making a 3D model, not bringing peace to The Realm.
Thanks :) yes I think I went a bit far with the music. Future ones will be music free :)
So I had the same crash as you, and managed to get enough debug data to make a good report on it. It's been found and will be fixed - though who knows when the next release will be, for people who don't compile it themselves. github.com/cnr-isti-vclab/meshlab/issues/500
specs?
If you mean specs of the PC I used, I think is was win7, i core5 3ghz, 12Gb ram, geforce 1050
Your forgot to tell that we should have MySQL preinstalled ;p
oh I did not know that, thanks!
猴赛雷啊
Loved your tutorial but the music made me want to sleep hard
lol, thanks :) I think i'll stop adding music in future.
I vote for no music! The video is really helpful - thanks! But the music isn't just unnecessary, it's a real distraction and at times swells to challenge hearing clearly what you're saying. Whatever music you put on, what type of percentage of viewers would like it and want it there? Very few I expect. And if I want some background music while following along, I've plenty of my own to choose from - and it will keep playing nicely regardless of going back over bits, pausing, and changing speed, etc!
Good to know, I shall no longer put music on my tutorials, cheers.
So fucking frustrating not knowing what you are doing with your shortcuts
Please remove the background sound. It can drive many people crazy!
I was wondering if people minded the music, I shall stop putting it on from now on.