IVML  
  about | r&d | publications | courses | people | links
   

Y. Xirouhakis, G.Votsis, and A. Delopoulos
Estimation of 3D Motion and Structure of Human Faces
Proc. of European Robotics, Intelligent Systems and Control Conference (EURISCON), Athens Greece
ABSTRACT
The extraction of motion and shape information of three dimensional objects from video sequences emerges in various applications especially within the framework of the MPEG-4 and MPEG-7 standards. Particular attention has been given to this problem within the scope of model-based coding and knowledge-based 3D modeling. In this chapter, a novel algorithm is proposed for the 3D reconstruction of a human face from 2D projections. The obtained results can contribute to several elds with an emphasis on 3D modeling and characterization of human faces. The problem of modeling human faces from their projections has been tackled before by several authors such as [1,2]. However, the accuracy of the obtained results is still far from the desirable. In [1] an algorithm is provided for the estimation of 3D motion and structure from two orthographic projections (two frames) having available a relatively accurate initial guess of the object structure, for example a generic wireframe model for the case of human faces. In [2] a modi cation to the above algorithm is proposed in order to increase its tolerance to errors in the initial guess. In the aforementioned approaches it is assumed that the human face is a gener- ally rigid object. Local deformations of the face, such as mouth and eye movement, are generally tackled separately as local motions. This constraint is also set in this work, i.e. the human head is assumed to be globally rigid. Similarly to other approaches, it is assumed that there is no camera movement and that the video is obtained under orthographic projection. The latter, as it can be seen in [6], yields satisfactory results and signi cantly reduces computational cost. In the present work, the algorithm presented in [5] is utilized in order to obtain an accurate solution for the object's 3D motion and structure (without any initial guess for the object's 3D structure) employing three orthographic projections. The algorithm exploits the information provided by the fact that the object is a human head, in the sense that keypoints and keylines are extracted increasing the accuracy of the 2D motion estimates. As it will be quoted in the sequel, accuracy in the 2D motion estimates plays a signi cant role to the accuracy of depth estimates.
01 January , 1998
Y. Xirouhakis, G.Votsis, and A. Delopoulos, "Estimation of 3D Motion and Structure of Human Faces", Proc. of European Robotics, Intelligent Systems and Control Conference (EURISCON), Athens Greece
[ save PDF] [ BibTex] [ Print] [ Back]

© 00 The Image, Video and Multimedia Systems Laboratory - v1.12