Facial Model Fitting Based on Perturbation Learning and It's Evaluation on Challenging Real-World Diversities Images

We present a robust and efficient framework for facial shape model fitting. Traditional model fitting approaches are sensitive to noise resulting from scene variations due to lighting, facial expressions, poses, etc., and tend to spend substantial computational effort due to heuristic searching algorithms. Our work distinguishes itself from conventional approaches by employing (a) non-uniform sampling features unified by the shape model that affords robustness, and (b) regression analysis between observed features and underlying shape parameters that allow for efficient model update. We demonstrate the effectiveness of our framework by evaluating its performance on several new and existing datasets including challenging real-world diversities. Significantly higher localization accuracy and speedup factors of 15 have been observed comparing with the traditional approach.