3D scanner to produce model and drawing - LARGE Scale, though
June 28, 2022 12:34 PM   Subscribe

How do I produce a series of drawings or 3d CAD compatible models on the cheap - is it possible?

So I have a large area that has some structure attached to it (a track and some kerbing next to it) and I need to get the actual info of how big and the realistic shape of it to a guy that is producing a computer model for me of the entire facility. I see Lidar scanners are real and seem to work and are possible on ipads and iphone and stuff but it is overwhelming what they actually are capable of, so I figured I'd ask you lot. It can be done in sections but we are likely looking at several areas of a maximum size of 100 feet x 40 feet. It *looks* like I can do that with an iPad/phone and it will give me accurate enough info that way, but they all seem to be using small object examples. Does it produce larger scale options?

Is this something that can be done? My head melted looking at all the examples and they are all linked to (at best) floor plans that look pretty simple compared to what I'd need in terms of complex shapes (curved kerbs with slope up, down and a flat section on each for example). I'd need it to be accurate to within 0.5" (probably be ok), but I'd need to produce a lot of sections of the track/kerb so I am not sure if there is some limit to how far I can move the ipad/iphone before it trips out. Is there a max size for these things? Any ideas? Is there a cheap service I can use to get it scanned?
posted by Brockles to Technology (8 answers total) 2 users marked this as a favorite
 
I've seen Corridor Digital use the iPhone Pros (with LIDAR) and PolyCam to scan the interior of an old mine shaft, so it's possible.
posted by kschang at 1:34 PM on June 28, 2022


It's not entirely clear from your question what you need to do with this model, or how critical the accuracy is to your organization. You need to "produce a lot of sections of track", does that mean you're going manufacture something based on this, like a train track, or... ?

My amateur understanding is that accuracy to within 0.5" will not be possible with consumer hardware like iPads, and that you'll want to bring in some professionals. Does the person building your 3D model have any guidance here?

You should look into Matterport. A few venues I know have had good experiences with them recently. You can choose to scan your facility yourself, or they can connect you with a service provider to do it.
posted by hovey at 3:59 PM on June 28, 2022


Response by poster: Sorry, it's a full size race car track with kerbing at the edges to replicate in a simulation software. If that helps.
posted by Brockles at 5:09 PM on June 28, 2022


I'm a researcher in computer vision, although specializing more in the deep learning side, but I have a lot of experience in this field. Building a 3D model of a large structure is a challenging problem both because of the physical scale, but also the sheer quantity of data you'll need to model it at a reasonable accuracy. You, broadly speaking, have two options:

Commercial LIDAR is great but is expensive and requires experts to gather the data and then turn it into a useful model. If you have 5 to 6 figures (USD) to spend, you will get a model accurate to the millimeter. You will not be able to use consumer-grade LIDAR (iPhone sensors, etc.) to generate models of anything larger than a few 10s of cm (a meter maybe?) on a side.

There is also photogrammetry -- building a 3D model from photos. This approach has historically been even more expensive than LIDAR because of the extremely difficult math behind it. However! There are some new techniques from the deep learning world that may work for you. One thing that would be easy (and cheap) enough to try would be some kind of neural radiance field latent geometry reconstruction.

Basically:
1. Take lots (hundreds? thousands?) of photos of the structure from every angle you can think of.
2. Use some kind of global SfM code (like COLMAP) to calculate the relative poses that all these photos were taken from.
3. Use the photos and COLMAP data and something like https://github.com/NVlabs/instant-ngp to generate a NeRF of your structure. Marching cubes will turn this latent geometry into a triangle mesh.

However, you still won't have scale (you can't recover scale from monocular images). So you'd have to physically measure some part of the structure and use that information to scale up the whole model. This is as easy as loading the triangle mesh into Blender and scaling by some number (length of thing in real world divided by length of thing in your model).

Some notes: the quality of your output is entirely dependent on the quality of your input. Use the best camera you have, and shoot high resolution, well-lit photos. This is the most time consuming part. COLMAP will run for hours (maybe many hours depending on how many photos you use). The NeRF code I linked above will run in seconds, so it's not the bottleneck. You'll want to use the most powerful (most memory, best GPU) computer you can find. COLMAP in particular takes a lot of memory, esp with lots of large photos. And your GPU needs to be from NVIDIA, preferably a 3090 but you might be able to get away with something lesser.
posted by riotnrrd at 5:31 PM on June 28, 2022 [1 favorite]


What’s your budget? I used to do this with a DJI Phantom 4 quadcopter and some very expensive photogrammetry software intended for engineering (Bentley ContextCapture). We had pretty darn good georeferenced point clouds and models by the time CC was done churning on a few hundred GPS-georeferenced photos on a grid , but ContextCapture is $$$$$. There are cheap and free alternatives but you’ll have to work harder to get to real-world scale. (One easy trick: capture an object of known scale, like a 2m X 2m checkerboard or something, and scale that to size in your 3d post processing tool)
posted by Alterscape at 6:48 PM on June 28, 2022


If this is for a simulation, I believe I read somewhere that Asseto Corsa or one of the similar caliber sims actually laser scanned the tracks.
posted by kschang at 8:32 PM on June 28, 2022


You don't even need a "3D" camera. You'll want a hefty PC (there might be a way to offload the processing online) instead.

Photogrammetry.

Agrisoft Metashape and RealityCapture if you're not terribly fussed about accuracy.

In practice, you might want to model "chunks" of the area individually, then stitch them together in the end. This gives you more flexibility on detail/ #polygons vs processing, for different use cases. There are different pricing models, RC has a not unreasonable pay-per-#polies-processed.
posted by porpoise at 8:41 PM on June 28, 2022 [1 favorite]


Land-surveyors will be able to do this to the accuracy that you need.

Where I live, in Scandinavia, scanning by Land-surveyors costs $150 - $200 per hour, for scanning and data processing.
posted by DelusionsofGrandeur at 12:27 AM on June 29, 2022


« Older Origin, age, quality of this rug?   |   What type of things does my equivalent in Spain... Newer »

You are not logged in, either login or create an account to post comments