"Predicting 3D People from 2D Pictures" co-authored by Ph.D. candidate Leonid Sigal and Professor Michael Black, won the best paper award at the Forth International Conference on Articulated Motion and Deformable Objects (AMDO-e 2006) in Mallorca, Spain.
Sigal's research focuses on building vision systems that can detect and track people in images and video. Such vision systems have many applications in the entertainment industry, rehabilitation medicine, surveillance and robotics. Black and Sigal’s latest work uses statistical models and a novel hierarchical probabilistic framework to infer the 3-dimensional pose of a person from a single image. This is a particularly challenging computational problem given the flexibility of the human body and the inherent ambiguity involved in computing a 3-dimensional articulated model from a 2-dimensional image.
This new framework breaks the problem down into a hierarchy of subproblems. First they detect the location of possible body parts in the 2D image and then these are combined into a 2-dimensional "cardboard person" model using a probabilistic graphical model. Machine learning methods are then used to predict 3D body poses from the 2D model. The result is one of the first complete systems for estimating 3D body pose directly, and automatically, from images.
The paper is available at: http://www.cs.brown.edu/people/ls/Publications/amdo2006sigal.pdf
