Which 3D scanning apps/services/projects/other or combinations of those would you recommend for producing high fidelity, 3D models of a real car?
Output format can be typical 3D engine format, e.g .obj .fbx .glb
Thanks in advance!
• Scaniverse - https://apps.apple.com/us/app/scaniverse-3d-scanner/id154143...
• Polycam - https://apps.apple.com/us/app/polycam-3d-scanner-lidar-360/i...
• Luma3D - https://apps.apple.com/us/app/luma-3d-capture/id1615849914
Higher quality (potentially), more time/effort (definitely)? Dedicated camera + desktop apps:
• RealityCapture (free) - https://www.capturingreality.com
• Metashape (paid) - https://www.agisoft.com/
My advice is to nail down what you're trying to do, end to end, which will help you narrow the field of what to use.
One of the problems you'll have is photoscanning does not do well with shiny surfaces, as the highlights move and confuse the reconstruction. If this were a model or maquette or something, they would probably spray paint it matte before scanning, which I doubt you want to do with a real car.
If you need high quality, honestly I'd just hire a hard surface modeller. There are people out there who make cars all day every day. If you just need low quality, Unreal's RealityScan app, Polycam, or AliceVision are good places to start.
^ uses OpenMVG: https://github.com/openMVG/openMVG
Colmap: https://github.com/colmap/colmap
Opensfm: https://github.com/mapillary/OpenSfM
Meshlab: https://www.meshlab.net/ (open source)
I also hear meshroom is popular but I haven't used it.
The implications of these requirements (ie car _in specific location_) is not obvious to me. What are they planning to do with the model?
Have they customized the frame and / or paintjob or is it a typical production model?
In the latter case it would likely be easier to source a 3d model rather than scan it.
If it’s some “playing around” Luma is pretty good nerf based scanner.
Without nerf you might have a hell of a problem with the specular surfaces.
If it’s production use with a budget I would recommend subcontracting the scanning - there are parties professionally doing this.
Photogrammetry relies on unique points on an object. Imagine that the software doesn’t have any information besides what’s visible on the photos. Basically it tries to match the same points on different photos and is then able to reconstruct the object by estimation.
This works great for objects which got a unique texture or many unique points. Like a newspaper, which has a lot of non-repeating text and thus a non-repeating texture. This doesn’t word great for single color objects, as the software can’t distinguish different points on the object.
This does also work on matte,opaque objects, as reflections will show the surroundings and may alter the color.
Now combine the above: cars are usually single color, shiny and have very little corners and edges. This is photogrammetries worst nightmare. Even if you coat the whole car in a single even layer of something, then it’s still single color…
What works (but it’s a painful task!) is to sprinkle colored matte dots onto the car, preferably in a random pattern. Then you take houndreds of photos of the car. It’ll take a long time and it must be done in an even lit place. So no direct sunlight but a cloudy day or inside a building with a lot of indirect lighting.
In short: Don’t try it and buy a scanner like Einstar instead. Renting a more expensive scanner is also an option. You’ll still need to coat the car, but it’s faster and less painful.
The downsides of the Vega I mostly didn't see mentioned on YT:
- Required an account to use
- Required WIFI to setup (don't need WIFI past that except when updating)
- Required an app to get scan files off it, but will work for your OBJ use case
Battery life has been a few scans. Do the processing later and get the scans up front.
(To be clear I love my Vega. Game changer. Just a bit heavy handed if you're not making car body parts)
For ~$1500 you can have a handheld 3d scanner. Their marketing materials show a scanned Hummer hood and a car door. I have never used it though so I am not endorsing it.
https://store.creality.com/products/cr-scan-raptor-3d-scanne...
Obviously there are usecases, but man are they a pain.
I have suggestions for getting visual fidelity but if you’re looking for an explicit 3D model maybe that’s not what you’re going for. (There are new approaches like NeRFs and Gaussian splats that don’t use traditional 3D formats, but look truly great.)
And good luck, whatever direction you take this!
The more accurate you want to get the more money you will generally have to spend.
And oh, you aren't likely to get a clean 3D model from the get go, more likely is that you are going to generate a point cloud that has to be edited to get a usable 3D model.