Face alignment is a fundamental yet challenging task for various facial analysis applications. Existing video-based face alignment (VFA) methods often neglect the consistency of facial geometries and textures across video sequences, thereby limiting their ability to handle accurate and stable face alignment. This paper presents a robust and highly accurate 3D Morphable Model (3DMM)-based VFA approach that leverages a novel texture generation method and a self-refined procedure for face alignment. By employing a differentiable rendering technique and a self-refined optimization method, we iteratively fine-tune facial geometries, textures, and poses. Experimental results demonstrate that our approach outperforms existing state-of-the-art methods in both accuracy and temporal stability.
ā±Coming soon