Face Alignment for Robust to Background and Occlusion Using Variation of 3D Reference Shape

Abstract

A Constrained Local Models is face tracking method based on variation of shape in the Allowable Shape Domain. The CLMs perform optimize strategy in patch. It has problem because occlusion and background. In this paper, variation of shape is constrained for solving this problem.The generating of shape allows only shape adjusted translation, rotation, scale. The proposed method prevent that the shape generate to abnormal forms. As a result, the alignment error is about 4.2 pixels in environment has complex background. Also, when part of the face occluded, the alignment error is about 18.1pixels.