Researchers at Stanford University like Andrew Ng, junior professor of computer sciences and his doctoral student Ashutosh Saxena developed a web service called Make3D. It can convert two-dimensional images into three dimensional models. So it is possible to get a more realistic impression of a still image.

How does the algorithm work?

To calculate the three-dimensional model machine learning is used. To teach the algorithm the relations between depth, orientation and position in a picture two dimensional images were used combined with 3D-data extraacted from 3D-Laser scanners. So the algorithm could get a relation between color, texture and size of individual areas of the picture and convert them into depth-values.

The photo will be covered initially into very small segments and for each of them the 3D-position, orientation and the angle will be calculated.

Historic

The attempt to create 3D models of pictures is not new. At the time models were synthesized from a variety of images by analyze them and put them together in a scene. The advantage of Make3D is of course that only one image is required to create a model. Using multiple images can improve and extend the scene known from the Microsoftapplication Photosynth.

A similar approach as Make3D already existed in May 2007 with the website Fotowoosh. Their algorithm had the problem that it could only handle surfaces vertically or horizontally.

Application areas for the algorithm

The algorithm can for example be used to create environments for computer games or create a 3D visualization of movies to broadcast in 3D cinemas.