Stabilized Temporal 3D Face Alignment Using Landmark Displacement Learning

One of the most crucial aspects 3D facial models is reconstruction. However, it unclear if face shape distortion caused by identity or expression when morphable model (3DMM) fitted into largely expressive faces. In order to overcome problem, we introduce neural networks reconstruct stable and precise faces in time. The reconstruction network extracts 3DMM parameters from video sequences represent Meanwhile, our displacement learn changes landmarks. particular, identity, expression, temporal cues, respectively. proposed alignment exhibits reliable performance reconstructing static dynamic leveraging these networks. 300 Videos Wild (300VW) dataset utilized for qualitative quantitative evaluations confirm effectiveness method. results demonstrate considerable advantages method sequences.