From Video to 3D Models


Australian Centre for Visual Technologies | VideoTrace: Rapid interactive scene modelling from video

The Video Trace system isn’t completely automatic. It requires a modeler to mark the boundaries of surfaces in at least a few frames of video. But it does seem extremely fast and well-integrated.

This isn’t quite ready to turn, say, Google’s Street View photos or video into 3D models just yet, but it’s close. Once the system can infer surfaces without human intervention, perhaps only requiring people to go in and fix mistakes, it might be ready to scale up to cities. Right now, it would still be too labor intensive. But it’s impressive nonetheless.

I tried linking the video here, but it’s being difficult. Click the link above for an interesting demo of their approach.

  1. #1 by Daniel on September 30, 2007 - 9:36 pm

    The low-latency modeling stuff is probably the ‘way to go’, I think, for automating 3D output — but until a video camera exists that can give accurate angle calculations from its location, it’ll be difficult to ‘video-reg’ each point.

    We were working quite a bit toward automatic 3D generation from 2D imagery — but the problem that exists there, is you’d need imagery that’s off-nadir. Pictometry, for example, is an ideal imagery source — but because it’s high-latency, it will never prove efficient as video for that purpose.

    Or so, that’s my take on it.

  2. #2 by Niraj on October 3, 2007 - 9:36 pm

    That’s a very slick tool…the video’s up on YouTube as well if you’re trying to embed it:

  3. #3 by Some scummy person on October 6, 2007 - 9:29 am

    Daniel – why wouldn’t a pair of video camera suffice? Or what about that idea I saw recently of including a mirror at one side of your field of view, so that you can see the same image from two offsets in one frame?

(will not be published)