I recommend using your smartphone for taking photos. Like the S22 for example. But before you start, turn on any metadata options you can find. It will write every sensorvalue including tilt, position, roll etc etc into the image metadata and photogrammetry tools can use those very efficiently!
@@rafiffadhil5872 It doesn't NEED the orientation as meta data. It just speeds up the calculation if it does. You can throw different cameras in the software (like mixing drone shots with regular shots) and the software will do just fine.
I used a similar workflow years ago to create some custom spring seat adapters for my Volvo. You have shown me a lot of little tips and tricks that would’ve made the process go A LOT more smoothly. Nice work!
Wow, this process has not improved at all in like 10 years. The results are way better now, tho. Thanks for the video, very clear instructions and tips!
@@horrorhotel46290 I think he means the laborious steps involved are the same as ever (no advances in automation), but the 3D reconstruction you get at the end is better. I don’t know whether that’s true, but it’s not contradictory.
hahah That’s pretty wild. I did that on one of my other channels. Just had some mobile game footage from years ago and posted a flight training video. (with a real airplane, not a video game) Just wanted to backup the files somewhere else besides the phone because it was overheating and the battery was expanding in thickness. Didn’t want to have to wait until I could order a new battery to get to the video files, in case the battery completely failed or even burst open. Ended up uploading public by accident and didn’t realize it till I got home hours later. Some of my old game buddys saw it along with a “face reveal” lol
Thank you for starting me down the path. I already have a 3d scanner, to use in a hobby / part reconstruction / replication of "antique" or out of production auto parts. For the last while I've had the scanner, I've been seeking an open-source focus on the toolchain and workflow. Thus, for me, I get to sidestep Meshroom for the moment, but everything in the MeshLab and FreeCAD portions still applies.
is it possible to use this process to scan the inside of something? I want to scan the inside of my truck, so that I can get some dimensions for fabricating a headliner. There are curves there that would be difficult to model in a traditional way.
I've been looking for an open source 3D scanning tool to add to my FreeCAD and 3D printing tool chain. I need to make some parts that mate to someone else's fairly complex parts and I don't have their models. I'm a bit put off by the CUDA and nVidia requirements, but photogrammetry is definitely getting better.
Using FOSS: Yay! And an Nvidia card.......... Oh.... Us FOSS users own ATI graphics cards since Nvidia refused to participate in FOSS for so long. Pity, was very interested in this video.
now this is a interesting topic... you could basically steal the blueprint of anything worth getting machined for example premium bike-frames or other stuff you can 3d-print or CNC it
I know two ways around that issue. Option A: Mount your object on a pole at a height that allows you to take pictures from above and below. Option B: Perform two scans, one with the object flipped upside down, and merge them digitally. Meshlab has the Alignment tool for that purpose.
The main idea of 3D scanning is that you don't need to do yourself 3D modelling. As long as it requires doing Topology manually after scanning it's not working as intended. You could 3D model this RC-Car by hand faster than you 3D scan it and then model manually
Not much. You'll still want to get pictures evenly around the object. Professionals use a drone for that. If that's not an option, take pictures close and further away from the house. Then, depending on your roof you might not get good reconstruction there. For scaling you could use the length of a wall for example.
A very fascinating thing. Tried myself 30 hrs ago. My thing is still rendering. I guess 70 photos with 108mp is overkill. Has Ryzen 5 1600x, 16gb of DDR4 RAM and Gtx 1060 6 gb to process it.
The number of pictures is good but 108mp is a lot ;) Are you stuck at the depth map node? That one took the longest for me. In the depth map node you can set a downscale factor. Try 8/16, the default is 2. On my system 100 images at 18mp took 80 minutes (ryzen 3800xt, 32gb ddr4, 1080TI).
@@DanielJutz It was just the first step. It took approx. 65 minutes to process one photo. So at 35 hrs it was around 30 photos out of 68. So i went to make another set of photos xD
maybe try using a better camera. Picture resolution isn't a good value to judge camera quality, usually they aren't good sensors if they need to market them with 108mp
Algorithmen often work better with less megapixels. Tesla's FSD example work I believe with a 3 or 5 Megapixel cameras. But sure it's not the same for fast real time 3d Images or high quality 3d files
Why do they use a comma in that case? Because they were already using the period as a thousands separator, of course! :) This sort of thing is quite like the idea of having traffic drive on the left or right hand side of the road as a convention. To be pedantic, it's called a radix point in general, in case you're not using a period, or a comma ... or not using a decimal number system. This issue is a problem of bad localization in MeshLab, and it ought to be raised as a bug with that project so this sort of advice to use a foreign choice of radix point symbol isn't needed anymore.
Yes that is possible. For it to work I had to set "Minimal 2D Motion" in the "FeatureMatching" Node. I used a value of 5 but you might have to experiment.
Like handwritten or printed? I feel like tesseract should be able to do it if it's printed. If memory serves there is a language pack for math symbols.
@@oliverer3 Yes, I have tried Tesseract as well as numerous open-source tools and private software like PDF elements. I have some old scanned books that are no longer available in my country, but they are low-quality images in the form of old PDFs. Unfortunately, I haven't come across a single AI tool that can accurately recognize the text and transcribe it directly into a format like Markdown or LaTeX, especially when it comes to mathematical formulas and other symbols.
@@Maisonier You could probably use a combination of tesseract and OpenCV to do it then but it would take a decent amount of manual effort to tune it. Plug and play tools are probably not there yet.
I was watching the video you referred to at the beginning when I saw yours. I clicked your video at the moment the same scene was showing in his, so I thought I was still on his! Very weird.
Draft Meshing is available which does not need a Nvidia GPU. While it computes faster, the results will always be worse. You can access that Pipeline from File > New Pipeline > Photogrammetry Draft. You will probably want to increase Describer Density and Quality to high or ultra in the Feature Extraction Node.
These days of Ai you can't even test any LLM without Nvidia, forget about cheap notebooks with Intel graphics or AMD trash (their chips are cheap but very bad in support and unreliable, Amd was found in planned obsolescence in graphics chips, so designed defects in Macbooks which massively dying after certain years, I have one of whose).
I recommend 3DF Zephyr (there is a free version to try it out) or better buy a good 3D Scanner such as Einscan SP or Einstar from Shining for slightly bigger parts.
@@theneonbop yeah but it’s around 70 pictures. Try it out. It’s enough for most projects. I bought the version for 300 pictures in the end because it works.
Seems way too much of a hassle compared to taking couple pictures from different axis and scaling picture to sketchplane and using that and simple measuring of target.
I don't know even any LLM project using anything other than Nvidia, so forget about Amd and never buy their trash - after they deliberately engineering the death of graphics chips in all Macbooks Pro from 2011, which I have, moreover I have two similar models Macbooks, one Nvidia, other Amd and last one died as predicted on forums, a mass time bomb. It's really hard to resurrect macbook after, you're left only with Intel graphics and half power slower notebook. Their chips dying because of overheating, ridiculous.
@@fontenbleauI don't know about the incident you refer to, but given it's an apple product I wouldn't be so quick to blame AMD. The LLM thing is because nvidia holds the rights to CUDA, not necessarily because they are inherently superior. Frankly I think in raw rasterization they offer better value than Nvidia in the most recent generation, but limited in gaming by ray tracing, and limited in everything else because of Cuda. They could make Cuda Red or something, but why would any developer support it when Cuda is already a standard with like 90% {made up number} market share and existing libraries. Their CPUs in recent years are genuinely so much better value and power efficienct than Intel's (and just better outright in gaming with 3D V-Cache) that I want their GPU department to start pulling their weight.
@@Bean-Time maybe if you would own such Apple model you will know, because apple forum filled about that, i've made the notorius recovery method by overheating it under blanket, but it worked for few days only and their current hardware even cheaply made. I have Mac mini from newest, it's horrible ventilated & overheating always.
I need LiDAR. I need to scan medium to large spaced and retain a high level of detail at mm or sub mm precision. And I need to be able to scan past 20 meters. The issue I have is that everything on the market that can do this starts at $50,000 usd. And goes well north of that. The fact that they can get away with charging this much money in 2023 is a joke. I’m can’t believe that someone hasn’t made something in the $1000-4000 range. And for the love of got please don’t say iPhones or iPads. I’m talking about real scanners.
I'd say it depends. While it does increase processing time, in my experience it also increases the likelihood of all camera positions being reconstructed successfully.
Better results are possible. Either by taking more pictures or by tweaking pipeline settings. You can find more information here: meshroom-manual.readthedocs.io/en/latest/faq/reconstruction-parameters/reconstruction-parameters.html . For those improvements you will have to accept longer processing times.
I recommend using your smartphone for taking photos. Like the S22 for example. But before you start, turn on any metadata options you can find. It will write every sensorvalue including tilt, position, roll etc etc into the image metadata and photogrammetry tools can use those very efficiently!
@@UCKY5 it's likely on by default, and unable to be disabled
@@justinlabarge8178 how do I check that my files have them?
@@UCKY5 The majority of smart phones don't allow you to turn it off besides GPS which is a good thing in this case.
How to record orientation of each photo? Idont get it. if it's drone, it separated to different file
@@rafiffadhil5872 It doesn't NEED the orientation as meta data. It just speeds up the calculation if it does.
You can throw different cameras in the software (like mixing drone shots with regular shots) and the software will do just fine.
I used a similar workflow years ago to create some custom spring seat adapters for my Volvo. You have shown me a lot of little tips and tricks that would’ve made the process go A LOT more smoothly. Nice work!
Wow, this process has not improved at all in like 10 years. The results are way better now, tho. Thanks for the video, very clear instructions and tips!
This is a contradictory statement
@@horrorhotel46290 I think he means the laborious steps involved are the same as ever (no advances in automation), but the 3D reconstruction you get at the end is better. I don’t know whether that’s true, but it’s not contradictory.
@@horrorhotel46290 a process is not a result..
Bro came back after 9years and just posted a super useful video. Very appreciated, thank you😁
hahah
That’s pretty wild. I did that on one of my other channels. Just had some mobile game footage from years ago and posted a flight training video. (with a real airplane, not a video game)
Just wanted to backup the files somewhere else besides the phone because it was overheating and the battery was expanding in thickness. Didn’t want to have to wait until I could order a new battery to get to the video files, in case the battery completely failed or even burst open. Ended up uploading public by accident and didn’t realize it till I got home hours later. Some of my old game buddys saw it along with a “face reveal” lol
You've made me very, VERY happy!! A nice concise way to do what I figured was possible but had no idea how!
Thank you for starting me down the path.
I already have a 3d scanner, to use in a hobby / part reconstruction / replication of "antique" or out of production auto parts. For the last while I've had the scanner, I've been seeking an open-source focus on the toolchain and workflow.
Thus, for me, I get to sidestep Meshroom for the moment, but everything in the MeshLab and FreeCAD portions still applies.
Meshroom is free and open source
I imagine he’s talking about time, not money
So, there is at least one person who knows how to use Meshlab. Impressive!
Cornflour mixed with IPA in a sprayer is another alternative to scanning spray.
Good video thanks. I had tried Meshroom before but struggled with it. This is a really concise and clear path through it. Good work!
Wow thanks. Scaling in MeshLab is so much easier
This was so great, took about 10 seconds to know I was going to subscribe! Thanks
Thanks for your time and trouble making this, much appreciated!
wow the scanning resoult is so good even without a 3d scanner crazy
fascinating and concise, very well put together video
Blender can take a video and turn it into individual images in a folder, FYI.
Thank you for a great walk-thru of the process.
Sadly the Cuda requirement is a dealbreaker :( Also, from an open source standpoint a solution for AMD cards would be lovely
Solutions for macOS would also be very welcome. Thanks!
Absolutely excellent video. I run the same software on Linux. Thank you for the video.
Great video with a lot of tips. Thanks!
Ahh, I was trying to do this photogrammetry using screenshots. Haha. Thank you for the video.
This is really well explained. Thank you! ❤🔥❤🔥
is it possible to use this process to scan the inside of something? I want to scan the inside of my truck, so that I can get some dimensions for fabricating a headliner. There are curves there that would be difficult to model in a traditional way.
Yes, it should work. Make sure to get photos from different positions in you truck.
Thats great, im planning on scanning all items in a computer museum to make a virtual tour
That sounds quite ambitious. Best of luck!
I've been looking for an open source 3D scanning tool to add to my FreeCAD and 3D printing tool chain. I need to make some parts that mate to someone else's fairly complex parts and I don't have their models. I'm a bit put off by the CUDA and nVidia requirements, but photogrammetry is definitely getting better.
I'm waiting for NeRF to get better. I think that will replace Photogrammetry and hopefully LIDAR also
This is EXACTLY what I need. Thank you!
great video, thanks for sharing your workflow in such detail
Very useful tutorial video made! Thx a lot 👍🏻
Using FOSS: Yay!
And an Nvidia card.......... Oh.... Us FOSS users own ATI graphics cards since Nvidia refused to participate in FOSS for so long. Pity, was very interested in this video.
I widh I could subscribe the second time
just for mentioning our bro Matt =)
Great Video, concise and really informative. Danke :)
11/10 explaination !
now this is a interesting topic... you could basically steal the blueprint of anything worth getting machined for example premium bike-frames or other stuff you can 3d-print or CNC it
Thanks for amazing tutorial! Question: I have a drone and want ro reconstruct a church, using this approach? Would it be possible?
From a technical standpoint it should work. Results will depend on your drone's camera. Always make sure you have permission to fly! ;)
Wouldn't it be easier when you start aligning meshes, to use a mirror-build from the middle, for certain shapes?
I hope something like this eventually comes to AMD GPUs. Haven't had an NVIDIA card since the 1050ti.
It will come to Intel GPUs.
time to join the winning side
1050ti is capable enough to run Meshroom
How do you take pictures of the bottom of the subject? There's a lot of detail underneath.
I know two ways around that issue. Option A: Mount your object on a pole at a height that allows you to take pictures from above and below. Option B: Perform two scans, one with the object flipped upside down, and merge them digitally. Meshlab has the Alignment tool for that purpose.
Thank you, This helped me very much.
The main idea of 3D scanning is that you don't need to do yourself 3D modelling. As long as it requires doing Topology manually after scanning it's not working as intended. You could 3D model this RC-Car by hand faster than you 3D scan it and then model manually
How would you change this workflow to make a 3D model of something big like a house or garden?
Not much. You'll still want to get pictures evenly around the object. Professionals use a drone for that. If that's not an option, take pictures close and further away from the house. Then, depending on your roof you might not get good reconstruction there. For scaling you could use the length of a wall for example.
For reflective surfaces you could creating a NERF rather than using photogrammetry.
A very fascinating thing. Tried myself 30 hrs ago. My thing is still rendering. I guess 70 photos with 108mp is overkill. Has Ryzen 5 1600x, 16gb of DDR4 RAM and Gtx 1060 6 gb to process it.
The number of pictures is good but 108mp is a lot ;) Are you stuck at the depth map node? That one took the longest for me. In the depth map node you can set a downscale factor. Try 8/16, the default is 2. On my system 100 images at 18mp took 80 minutes (ryzen 3800xt, 32gb ddr4, 1080TI).
@@DanielJutz It was just the first step. It took approx. 65 minutes to process one photo. So at 35 hrs it was around 30 photos out of 68. So i went to make another set of photos xD
108mp?!? What are you shooting with?
maybe try using a better camera. Picture resolution isn't a good value to judge camera quality, usually they aren't good sensors if they need to market them with 108mp
Algorithmen often work better with less megapixels. Tesla's FSD example work I believe with a 3 or 5 Megapixel cameras. But sure it's not the same for fast real time 3d Images or high quality 3d files
Try nerf it even work on reflective surface as well
Excellent video! What RC truck is that?
Thanks! An old nitro rc car converted to electric with a drill motor.
We can't use Radeon Graphic?
4:13 you mean use a decimal as a decimal? Why do Europeans use commas for decimal points?
Why do they use a comma in that case? Because they were already using the period as a thousands separator, of course! :)
This sort of thing is quite like the idea of having traffic drive on the left or right hand side of the road as a convention. To be pedantic, it's called a radix point in general, in case you're not using a period, or a comma ... or not using a decimal number system.
This issue is a problem of bad localization in MeshLab, and it ought to be raised as a bug with that project so this sort of advice to use a foreign choice of radix point symbol isn't needed anymore.
Awesome! Instant sub. ❤️
can we use photogrammetry with turntables instead of us rounding the object?
Yes that is possible. For it to work I had to set "Minimal 2D Motion" in the "FeatureMatching" Node. I used a value of 5 but you might have to experiment.
Blender can replace meshlab, and will give you undo support, just gotta use a different workflow.
What about taking a video and then splitting it up into a series of still images? Seems like one could get thousands of pics that way?
Photos have higher resolution
Thanks, very comprehensive
Nice video. Will this work if I record a full 360 video in 4K and export that as single images?
Maybe, Meshroom might not like it if you are in the frame. This discussion might be able to help you github.com/alicevision/meshroom/issues/526 .
@@DanielJutz thx.
I'm still waiting for a good OCR AI that can read old math books and transcribe all the formulas without any issues
Like handwritten or printed? I feel like tesseract should be able to do it if it's printed. If memory serves there is a language pack for math symbols.
@@oliverer3 Yes, I have tried Tesseract as well as numerous open-source tools and private software like PDF elements. I have some old scanned books that are no longer available in my country, but they are low-quality images in the form of old PDFs. Unfortunately, I haven't come across a single AI tool that can accurately recognize the text and transcribe it directly into a format like Markdown or LaTeX, especially when it comes to mathematical formulas and other symbols.
@@Maisonier You could probably use a combination of tesseract and OpenCV to do it then but it would take a decent amount of manual effort to tune it. Plug and play tools are probably not there yet.
I was watching the video you referred to at the beginning when I saw yours. I clicked your video at the moment the same scene was showing in his, so I thought I was still on his! Very weird.
Any options for AMD cards?
Sehr gutes Video!
Any way to do this without a monster graphics accelerator?
Yes, the exact same way with a cheaper GPU
Is nVidia strictly required?
Draft Meshing is available which does not need a Nvidia GPU. While it computes faster, the results will always be worse. You can access that Pipeline from File > New Pipeline > Photogrammetry Draft. You will probably want to increase Describer Density and Quality to high or ultra in the Feature Extraction Node.
These days of Ai you can't even test any LLM without Nvidia, forget about cheap notebooks with Intel graphics or AMD trash (their chips are cheap but very bad in support and unreliable, Amd was found in planned obsolescence in graphics chips, so designed defects in Macbooks which massively dying after certain years, I have one of whose).
Thanks for sharing! 🙂
I recommend 3DF Zephyr (there is a free version to try it out) or better buy a good 3D Scanner such as Einscan SP or Einstar from Shining for slightly bigger parts.
I also prefer 3df zephyr, is much much faster imo. Free version used.
the free version has a limit on the number of images you can use
@@theneonbop yeah but it’s around 70 pictures. Try it out. It’s enough for most projects.
I bought the version for 300 pictures in the end because it works.
why not just scan with an iphone with the lidar. does all the computing on the phone in the app and puts in scale and even color pretty accurately.
I got an error message for even thinking about doing this.
Fascinating!
Before I use something like meshlab I use blender which is way better for cleaning and also retopo
What is a much more powerful device. Calipers.
Seems way too much of a hassle compared to taking couple pictures from different axis and scaling picture to sketchplane and using that and simple measuring of target.
take a shot every time he says 'mesh'
AMD 6900xt?
i used to do this for fun a long time ago using "123d catch" and later "3DF Zephyr" 😅
"An nvidia gpu"
*pulls out a decidedly red card* 😂😂😂
0:20 isn't that an AMD Gpu?
It's a Nvidia 1080 Ti.
I don't know even any LLM project using anything other than Nvidia, so forget about Amd and never buy their trash - after they deliberately engineering the death of graphics chips in all Macbooks Pro from 2011, which I have, moreover I have two similar models Macbooks, one Nvidia, other Amd and last one died as predicted on forums, a mass time bomb. It's really hard to resurrect macbook after, you're left only with Intel graphics and half power slower notebook. Their chips dying because of overheating, ridiculous.
@@fontenbleauI don't know about the incident you refer to, but given it's an apple product I wouldn't be so quick to blame AMD. The LLM thing is because nvidia holds the rights to CUDA, not necessarily because they are inherently superior. Frankly I think in raw rasterization they offer better value than Nvidia in the most recent generation, but limited in gaming by ray tracing, and limited in everything else because of Cuda. They could make Cuda Red or something, but why would any developer support it when Cuda is already a standard with like 90% {made up number} market share and existing libraries. Their CPUs in recent years are genuinely so much better value and power efficienct than Intel's (and just better outright in gaming with 3D V-Cache) that I want their GPU department to start pulling their weight.
@@Bean-Time maybe if you would own such Apple model you will know, because apple forum filled about that, i've made the notorius recovery method by overheating it under blanket, but it worked for few days only and their current hardware even cheaply made. I have Mac mini from newest, it's horrible ventilated & overheating always.
Danke, aber es wirkt jetzt nicht besser als noch vor ~4 Jahren
I need LiDAR. I need to scan medium to large spaced and retain a high level of detail at mm or sub mm precision. And I need to be able to scan past 20 meters.
The issue I have is that everything on the market that can do this starts at $50,000 usd. And goes well north of that. The fact that they can get away with charging this much money in 2023 is a joke. I’m can’t believe that someone hasn’t made something in the $1000-4000 range. And for the love of got please don’t say iPhones or iPads. I’m talking about real scanners.
> Open Source Photogrammetry
> Uses an nvidia only tool on windows.
Yeah, right...
And on proprietary CPU, shame on him should be all open source to count!
You don't do your computer any favours including all that paving in the images
I'd say it depends. While it does increase processing time, in my experience it also increases the likelihood of all camera positions being reconstructed successfully.
scan with iphone 11 is better
i love the vid but the result is poor
Better results are possible. Either by taking more pictures or by tweaking pipeline settings. You can find more information here: meshroom-manual.readthedocs.io/en/latest/faq/reconstruction-parameters/reconstruction-parameters.html . For those improvements you will have to accept longer processing times.
@@DanielJutz but is it better than a manual topo ?
Ok I dont want to sound like a dick but how is requiring cuda open source ^^
Photogrametry sux!
You are really good at working with the software, but you not that good at explaining stuff.my brain is hurting
Maybe open source but still proprietary thanks to CUDA so unavailable for anyone not wanting to support the contender for the worst tech company.
but meshlab sucks soooo bad, sooooooo bad....
This is a research tool, not a production tool. So the guy spends a whole day making just one part. Big deal.
Great video, slick system, very informative, Thank You!!!