Disney finds new way to give movies depth
2D pic to 3D background tech came too late for 'The Lone Ranger'
Disney Research in Zürich is working on ways to improve how 2D composite images can be turned into 3D models.
The basic techniques are well known, but – according to the research the group presented at the recent SIGGRAPH conference – there are limitations. Lasers can capture data to present a 3D surface, but are blocked by obstacles; and in cluttered environments, blending large numbers of photographs to derive a 3D image is very processor-heavy.
To make processing manageable, Disney Research says in a (PDF) press release, “Most existing stereo reconstruction techniques have been tailored for resolutions of just 1 or 2 megapixels”.
Disney's work has concentrated on the algorithms that calculate the difference between 2D scenes to calculate depth estimates. Even though they say they can work at much higher resolution (the demonstration scenes were shot on a DSLR at 21 megapixels), and calculate depth at the individual pixel level, the image data “can be processed efficiently with a standard GPU”, they claim.
To overcome the processing problem, the researchers use the coherence in “massive light fields” as the basis of their algorithm (it would seem to Vulture South that detecting single-pixel coherence is a feat in itself, but we're happy to be corrected on that point). From there, they say, the algorithm:
- Computes depth estimates around object boundaries rather than interior regions – this is because object boundaries are where it's easiest to calculate differences;
- Then, “more homogeneous interior regions are … processed in a fine-to-coarse procedure rather than the standard coarse-to-fine approaches.”
The Disney research paper states that a hundred 3 megapixel images could be depth-processed in 162 seconds per view.
In spite of the levity in the headline, there is of course a serious point to all of this, dead centre in Disney's business: movie-making. The world of CGI is famous for its consumption of processor cycles, and this offers an efficient way to create city-scape 3D backdrops with a GPU and StreetView-style image captures. ®
Sponsored: DevOps and continuous delivery