Date of Award
2008
Degree Type
Thesis
Degree Name
Master of Science
Program
Computer Science
Supervisor
Dr. Yuri Boykov
Second Advisor
Dr. Olga Veksler
Abstract
Face tracking is a widely researched problem and many different approaches have been developed for this problem. One promising recent method is based on “Pictorial Structures” [16]. A pictorial structure is a collection of parts represented by a partspecific appearance model and a deformable model allowing different configurations of parts. Pairs of parts are connected via spring like connections. An energy function is formulated to measure the individual part matches (based on a appearance model) as well as how well the configuration of parts match a deformable spatial model. In Felzenszwalb and Huttenlocher’s work [16], the face model is two dimensional. In this thesis, we extend the pictorial structures face model to three dimensions. We acquire the depth information (the third dimension) from pairs of images obtained from a stereo cameras. By exploiting information gained from the third dimension, our algorithm achieves a greater robustness in face detection results. Results show that by incorporating depth, face detection results increase with an absolute amount by 4 to 13 percentage points (a relative amount by 14 to 25 percentage points), when compared to the results of Felzenszwalb and Huttenlocher [16].
Recommended Citation
Luong, Robert, "Pictorial Structures for 3D Face Detection" (2008). Digitized Theses. 4799.
https://ir.lib.uwo.ca/digitizedtheses/4799