Semi-Supervised Learning for Facial Landmarks with Confidence and Augmentation Sifting Mechanisms

Facial landmarks are important for various facial analysis tasks, including face recognition, age estimation, expression identification, medical image processing, and forensics. Influenced by the popularity of self-training in recent years, we propose a semi-supervised based human face landmark detection algorithm. First, we train a model with labeled data. Then, a huge amount of unlabeled data is fed into the model to generate pseudo labels. In order to filter out the pseudo labels with higher credibility, we propose a probabilistic model and determine how close the output feature distribution corresponding to the pseudo labels to the Gaussian distribution is. Then, the data with the pseudo labels are adopted to improve the performance. Moreover, different thresholds are applied for screening. Experiments show that, with the proposed semi-supervised based algorithm, the accuracy of landmark extraction can be improved.