Hi all, I am new to the blender community and am hoping that someone out there can help me in my search.
I am wondering why I can’t find a 3D modeler that can take any random number of photos and use a best guess algorithm to stitch the photos together automatically to create a 3D mesh? (A sort of poor man’s scanner with a bit more oomph)
I imagine such software would be capable of combining Booleans and displacement mappings and analyzing perspectival restrictions to yield a mesh.
This is the way that the mind processes the 3D world so why is there no software that can mimic this?
This would allow immeadiate input from an intuitive source and, to my mind, be a very flexible tool.
I have tried various programs: Imagemodeler 3, Canoma, Dsculptor among others but as I discovered they all have at least one fatal flaw in that they require the user to physically trace an outline around the desired model. This not only seems self defeating, time consuming and generally counter productive but also doesn’t seem to optimally use the computer’s capability to calculate. I haven’t even seen one of the aforementioned programs that allows you to use a ‘magic wand tool’ to create an outline!
Perhaps someone out there in the community knows of some program or other solution that I am, as yet, unaware of, if so I would be very grateful to hear about it.
I am also open to discussing the development of such a piece of software.
Please feel free to email me at [email protected]
Thanks
This would be really nice. But I have a few questions about theory. First off, we know the theory is possible because it’s what humans do. But I feel that our conciousness play a part in it. Because we can search our memory for recongizable objects, we can then use that to expect what the general shape, size, etc… of that object will be. We would then use color and highlights (because color alone can be deceiving) to determine the three-dimensionallness (i don’t think that’s a word) of the object. But a computer would have no past reference or a euclidean plane of geometry to call upon as we do (unless through AI). Although it could recongize primitives, the colors and highlights in photographs could be decieving and blur the definition between the primitives.
So I think there is no doubt you wouldn’t get perfect results sometimes, but the question is, how messed up would the imperfect results be? At least by making the user select the area, human conciousness and our ability of higher reconigition come into play.
Can we recreate the coucious thought process that would allow the computer to ‘learn’ what it is seeing, so that the rate of error would decrease with practice?
Hi acasto,
Seems you might need D Sculptor!
http://www.d-vw.com/dsculptor/about/about_d_sculptor.htm
It’s quite expensive but the full version 1 is on the cover disc of 3d World of july 2002 so you could order a back copy of the magazine.
It’s a nice program, you have to print out a sheet with calibration dots and photograph your object on top of the sheet. In each photograph you’ll have to tell the software where the dots are and ot makes a fully textured mesh. Not so nice, this kind of 3d scanning doesn’t take holes into account, so it’s only good for simple objects.
I thought it was a bit disappointing, but try it:)
Have fun,
WAVK
Actually, for 3D ‘scanning’, don’t computers do better at scanning real 3D object with lasers and camera than it would at converting a 2D image to 3D?
The ICARUS program deals with this as well, you can find an import script for the camera calibration in the python & plugins forum.
But ICARUS also isn’t automatic, you still need to do the modeling yourself.
Wavk - If you look at the original posting I mention D Sculptor and I was disappointed too.
Acasto - All the component parts of my theorectical program are already possible. It only requires someone to put it together and that someone is who I am looking for.
Also you would be right to suggest a scanner as I did in my original posting if I was only going to use ONE photo! On the other hand using multiple photos MUST contain enough information to create a 3D model. Also the 3D Modelling I propose becomes accessible to everyone when you find an alternative to a scanner.
Further, there is a new paradigm being proposed that reinterprets the nature of 3D object creation and offers a simple alternative to a complex problem. I am asking you think in a non-linear way, can you do that?
Eeshlo - thanks for the suggestion. I assume you don’t know of any other software that doesn’t require tracing or uses a ‘magic wand’ tool to reduce the banality of laborious tracing?
I invite anyone interested in working on this project to contact me directly at my email add. [email protected]
I understand that. But even if you used multiple photos, they would have to be the perfect photos. The contrast would have to be right, objects would all have to be perfectly discernable, etc… For instance, where I live, there is a bunch of rivers with large boulders. Suppose I put multiple photos from multiple angles in. The boulders don’t have a defined geometry, so unless the computer knows what side is what through some type of marking, it wouldn’t know the difference bettwen a new boulder, and the same one at a different angle.
But isn’t that how a computer thinks?
Acasto - you seem to be missing the point: No matter how many pictures you take of your river or from what angle or under what lighting conditions, there is only ONE possibile 3D construction that will make sense. In fact the more unique the object and the more detail and variations it has, the easier it will be for the computer to pick out exactly what is consistent in all scenerios, thereby yielding a more accurate model.
I am not claiming that it wll be a perfect 3D copy of what you see but if the software is written correctly it will be a very good representation of it. It is exactly this sort of calculation that computers do best.
BTW Computers DON’T think NON-linearly.
I get the point. But it can’t make any picture if it don’t know how to handle the material it is given. Say you put a few pics of the river in there, then a pic of a fish. Does it laugh and say “that pic dosn’t belong”, or dose it give it it’s best shot. I think this is possible, but it will need some type of intelligence. Also, if only take a handful of vertices out of place to make a scene all screwed up.
BTW, I meant comptuers DON’T think NON-linearly. You asked if I could think non-linearly. So I meant why think about a problem non-linerly that is going to have to be solved in a linear manner.
ok guys go take this banter off to an AI chat or something. That would have been my first reply. What you’re asking for is a human mind put to a computer, and to create something only with systematic means is, as of yet, impossible.
-RonC
If I understand correctly the program your looking for uses some sort of pattern recognition as well as photogrametry to calculate angles in 3 dimentions without the help of a CG artist. Of coarse don’t forget that the brain needs stereoscopic vision to do this accurately. I certainly can’t help you there however I do know of a “manual” photogrametry program (which you may already know of) called Photo Modeler. http://www.photomodeler.com/
There is a light (free) version and a pro version. It uses special algorithms to extrapolate 3d information from 2d images. Now the meshes produced by the program are rudimentary at best. Say about as good as an average 3d video game if you’ve mastered it. But It’s real value is as a tool for science because it can closely approximate the dimensions, angles, proportions, etc. of real objects in photographs.
And the meshes can be imported into Blender and other apps. if you save
them in VRML format.
There is only one way of doing this completely accurate. This is modeling it from scratch using a photo as reference. All of those programs have flaws.
The best one that tries to do this automattically is D-Sculpter. It is hundreds of times better than the others. If that didn’t work for you, nothing will. You could also try Photomodeler http://www.photomodeler.com/ which takes a bit extra work.
If anyone is really interested in this, do a search for ‘image based rendering’ and/or ‘image based modelling’. You will find plenty of links with research being done in this field, besides some simple demos here and there, I don’t know of any program that actually makes use of these concepts yet.
I just wanted to say thanks to the last few posting posters 
This is much more the response that I have been hoping to get.
If anyone does find a link to reasearch in this direction, I would love to hear about it. It means a great deal to me. Conversely if anyone wants to get together to discuss it in greater depth, I am open to this as well.
PS I live in Amsterdam. you can find my email posted on the first msg link