An Improved Quadric Error Metrics and Mesh Parameterization Based 3D Face Intrinsic Attributes Descriptor
Guo Zhe Zhang Yan-ning Lin Zeng-gang
Shaanxi Provincial Key Lab of Speech & Image Information Processing(SAIIP), School of Computer Science, Northwestern Polytechnical University, Xi'an 710129, China
Abstract：A novel 3D face intrinsic attributes descriptor is proposed in this paper. The presented descriptor is used to solve the feature inhomogeneous distribution and redundancy for high resolution 3D face model. In presented method, an improved Quadric Error Metrics (QEM) mesh decimation method is first developed based on semi-edge folding and adaptive cost weighting. Multi-resolution 2D intrinsic attributed image can be then obtained by homeomorphically mapping 3D decimated facial mesh into 2D plane with the highest attribute preserving based on feature restricted conformal isomorphic mapping. Consequently, 3D surface matching issue can be simplified to a 2D image matching issue by comparing the resulting 2D intrinsic attributed images, which are stable and robust to occlusion and noise. Experimental results on GavabDB show that presented method has the ability to represent intrinsic information of 3D face and achieve significant improvements on recognition accuracy compared with baselines.
2010, Vol. 32 
