New Scientist reports on new computer vision research that allows the learning of surface structure using two digital images taken from the same viewpoint but under different illumination. More specifically, one photo is taken with flash and another using only ambient light. The differences between the two images allowed researchers from the University of Manchester in the UK and Dolby Canada to extract albedo and depth information for the surface; such information is then used to render the surface from different viewpoints properly illuminated. Under the title of “surface depth hallucination,” the researchers presented their work at the recently held, well known and very prestigious SIGGRAPH conference.
The project's web site describes the method as follows.
Surface depth hallucination offers a simple fast way to acquire albedo and depth for textured surfaces that exhibit mostly Lambertian reflectance. We obtain depth estimates entirely in image space, and from a single view so there are no complications that arise from registering texture with the depth obtained. (source)
The following video from New Scientist explains the surface depth hallucination method in more detail and it includes examples of estimated surface models.
1 comments:
10:28 AM
This looks cool, but unfortunately only works under limited circumstances. There are plenty of situations which for example a robot in a domestic environment might encounter where the surfaces would be non-lambertian or the lighting non-diffuse.
Perhaps future planetary rovers could use this technique to generate 3D models of rocks.
Post a Comment