Examples of Automatic 3D reconstructions from photographs

paulbourke.net

31 points·dualogy··7 comments

Tangental random thought: I wonder if you could reconstruct a high-quality 3D concert from combining multiple pieces of low quality cell phone video taken by fans...

http://switchcam.com/ is starting down that path.

I looked at this briefly, and it would be extremely tough. Very few events have the kind of coverage (i.e., number of input videos) you would need to even get started on the problem. Plus, the conditions are just about the worst possible -- highly dynamic scene, low light, very unsteady cameras, almost all viewpoints in the same plane, low quality inputs, etc.

Note that the switchcam startup mentioned in the other comment is almost completely manual right now, and is more about assembling a "director's cut" version of a concert video, rather than creating 3d. This actually seems more useful to me (although of course more automation would be nice).

There was a computer vision paper several years back where they did something similar, using the audio to synchronize the various cameras. Much lower setting complexity and good lighting though, unlikely to work in a concert setting.

I've also gotten good results with ScannerKiller[1] from XYZRGB, for meshes where very high accuracy is needed/wanted.

[1] http://www.scannerkiller.com/

This looks like an incomplete page, or perhaps a work in progress. Which method was used for generating each mesh? How many input images were there? What do they look like?

What's the takeaway message?

I've used RunSFM (which uses bundler) to make 3D models from photos. It's cool, but you need to have lots of high quality photos for it to work.

I'd like to play with using a Kinect or simular 'depth camera', it might give better results.

Reply on news.ycombinator.com