2018/06/21 by Jie Cao, Cao, Jie, Yibo Hu +7 · 5 citations
Computer Science · Engineering · #Face recognition and analysis #3D Shape Modeling and Analysis #Face and Expression Recognition
paper · pdf · doi:10.48550/arxiv.1806.08472
Face frontalization refers to the process of synthesizing the frontal view of\na face from a given profile. Due to self-occlusion and appearance distortion in\nthe wild, it is extremely challenging to recover faithful results and preserve\ntexture details in a high-resolution. This paper proposes a High Fidelity Pose\nInvariant Model (HF-PIM) to produce photographic and identity-preserving\nresults. HF-PIM frontalizes the profiles through a novel texture warping\nprocedure and leverages a dense correspondence field to bind the 2D and 3D\nsurface spaces. We decompose the prerequisite of warping into dense\ncorrespondence field estimation and facial texture map recovering, which are\nboth well addressed by deep networks. Different from those reconstruction\nmethods relying on 3D data, we also propose Adversarial Residual Dictionary\nLearning (ARDL) to supervise facial texture map recovering with only monocular\nimages. Exhaustive experiments on both controlled and uncontrolled environments\ndemonstrate that the proposed method not only boosts the performance of\npose-invariant face recognition but also dramatically improves high-resolution\nfrontalization appearances.\n