AI 3D models from photo?

[Chris333]

I see a few places now claim to be able to do this, make a 3D file from photos. Just wondering if anyone here has played with it and how good are the results. Most of them want you to signup for a plan $$$ to play and I just don't want to waste money on crap.

This is one: https://www.app.kaedim3d.com/signup

[Chris333]

And checking into the plans. The cheapest plan is $150 a month.

[Erock482]

You may want to look into reality capture if you have the prototype in front of you or available to take pictures of. It can stitch together lots of photos to then generate a point cloud and finally a 3D model from that.

For example. something like this: https://www.reddit.com/r/3Dprinting/comments/rt2rcc/i_3d_scanned_this_collapsing_1890s_granary/

I believe one of the software options allows you to generate the render, and then only pay for what you use. which is based on the amount of points/images processed.

If you're trying to generate stuff from a single image. kadeim looks like it might work if you can cut out the background. But that is also going to depend how high quality your source image is.

[mramsey]

The technique is called photogrammetry, no need for AI.  There are several open source (free) tools available, although all have a learning curve.  The most well known are:

• Meshroom
• MicMac
• COLMAP

[Chris333]

I was mostly looking to just get overall dimensions of a single photo. If I was able to go there and take photos, I could just measure it.

[Scottl]

Photogrammetry requires at least two photographs of the same scene/object from different perspectives.  The most well known use of this technique is aerial photo analysis but the principles apply to any number of photos (and the algebra to any number of dimensions).  It is routinely used for measurement of features and dimensions but requires a means for absolute scale such as a object of known size in the image.  Lots of packages including the freee one we use for UAV imagery Cloud Compare https://www.danielgm.net/cc/.  There are likely 3D scanning versions that optimize for objects that you can take a lot of images, but full coverage and shadowed areas are problems with the technique.

[Missaberoad]

It's Insane how fast this is all coming. Every day I'm amazed at something we can do now!

"Tea, Earl Grey, Hot."

[learmoia]

"Tea, Earl Grey, Hot."

There is probably some voice activated beverage machine that can do that today.  Though it wouldn't generate the cup out of nothing....

(Is there a scientific explanation of how a true replicator could work?)

It's Insane how fast this is all coming. Every day I'm amazed at something we can do now!

[Cajonpassfan]

You may want to look into reality capture if you have the prototype in front of you or available to take pictures of. It can stitch together lots of photos to then generate a point cloud and finally a 3D model from that.

For example. something like this: https://www.reddit.com/r/3Dprinting/comments/rt2rcc/i_3d_scanned_this_collapsing_1890s_granary/

I believe one of the software options allows you to generate the render, and then only pay for what you use. which is based on the amount of points/images processed.

If you're trying to generate stuff from a single image. kadeim looks like it might work if you can cut out the background. But that is also going to depend how high quality your source image is.

I don't know the scale of that granary model, but its three dimensional depth is phenomenal.
Try to build something like it from scratch
Otto K.

[Chris333]

This is what I was dreaming of. I have this photo:


That I fixed with AI:


And here is the spec sheet (seems they used this very locomotive in their catalog).


I know the locomotive had 9x14 cylinders so I can assume it has 28" drivers with a 5'10" WB and is 17' long overall, but I would love AI to spit out a 3D file to see if it matches the "guesses" that I drew up.

[Scottl]

One image like this won't go anywhere with photogrammetry, but is useful for some scaling measures.

Spit balling.  Locomotives have common structures (boiler, cab, domes, wheels, drivers, etc.) and symmetry that the AI can infer the 3D shape from other pictures, and then combine it scaled to this particular photo.   I'm sure it can be done as AI will drawn on a lot of data that is available, not just the picture in question. No doubt there are a lot of pictures out there of similar smaller Porters. 

I don't know how, but I can envision how it would be done.

[mramsey]

So, I plug the words "porter 0-6-0" into Google, and in short order I come across this thread.  My comment: No alleged "AI" tool is going to produce a more accurate 3D model of that locomotive from that single photo than you have already managed to put together.  There is almost certainly insufficient properly labeled training data incorporated into current generative models to come even close.  I know a PhD student who is in the midst of training an ML model for a tool that accepts hand-drawn sketches and a few dimensions as inputs and generates simplified 3D building envelope models in a limited set of shapes.  That is not all that far from the current state of the art.

By the way, the photo shows barrel distortion from the wide angle lens used to get the desired closeup detail.  That's why you can see the front of the cab, but not the front of the smokebox door.  There are tools to fix that to some extent (like PhotoShop and similar open source tools), possibly allowing you to measure more accurate dimensions on the image.

Perhaps you can find more prototype documentation?  What books, libraries, and/or museums might have photographs or erection drawings of H. K. Porter locomotives, either this one or others of similar vintage?  Does anyone have business documents (and perhaps photos) belonging to the G. M. & J or G. B Sandomen Co.?

[Peavine]

There’s a Porter 0-4-0 at the Sharlot Hall Museum in Prescott, Az.  Not sure if that helps, but it’s there…

[Chris333]

The 3D print on that page was just a quickie that was push/pulled from another locomotive I drew up. The tank is too long for sure, but it was all pretty much a guess. I have this outline drawn up now:


Knowing from the photo caption when it was built I could look it up in my Porter book. All that lets me know is it was 36" gauge and had 9x14 cylinders. The page from the old Porter catalog tells me what the driver size (and WB and overall length) is on this model with 9x14 cylinders.

The old Porter records are somewhere in Canada and the few people who knew about it, couldn't tell me how to obtain or search for erection cards.

As for Porter 0-4-0's. I have a bunch of scale drawing of those.

[Chris333]

Somewhere in Canada =  MUSEUM OF SCIENCE & TECHNOLOGY, Ottawa, Ontario., Canada

If anyone knows a contact there the construction number of the locomotive is #347 for the Green Marcus and Johnson quarry in New Castle, PA. I'll pay for the info!