Cascade Regression-Based Face Frontalization for Dynamic Facial Expression Analysis

Abstract Facial expression recognition has seen rapid development in recent years due to its wide range of applications such as human–computer interaction, health care, and social robots. Although significant progress been made this field, it is still challenging recognize facial expressions with occlusions large head-poses. To address these issues, paper presents a cascade regression-based face frontalization (CRFF) method, which aims immediately reconstruct clean, frontal expression-aware given an in-the-wild image. In the first stage, shape predicted by developing regression model learn pairwise spatial relation between non-frontal face-shape counterpart. Unlike most existing prediction methods that used single-step regression, multi-step regressor gradually aligns view. We employ several different regressors make ensemble decision boost performance. For texture reconstruction, active appearance instantiation employed warp input generate clean face. remove occlusions, we train generative on manually selected clean-face sets, ensures generating output regardless whether involves or not. reconstruction are computational expensive, proposed method works real time, so suitable for dynamic analysis expression. The experimental validation shows ensembling improved accuracy average 5% achieved superior performance both static over state-of-the-art approaches. results demonstrate expression-preserving frontalization, de-occlusion recognition.