Cross-Dataset Pose Estimation of Faces In The Wild

Abstract

In this paper, we propose a domain generalization method for cross-dataset pose estimation of faces captured in wild conditions. Conventional methods mainly devote efforts on extracting discriminative features to reason the three-dimension pose. Due to the large distribution discrepancies between widely-used synthetic training and real-world testing data, it is challenging to seek a domain-generalized feature space especially for the new test samples in real-world applications. To alleviate the influence of dataset bias, our model aims to learn the domain-invariant features across different domains. In detail, a carefully-designed domain discriminator is plugged to the features extracted from different domains, meanwhile the feature encoder is trained to enforce features from different domains confused by game-theorem iterations. With the adversarial manner, our model learns a generalized pose-relevant feature space shared across different domains. Extensive experimental results on the standard benchmark under the cross-dataset setting indicate the superiority of our method in comparisons with most state of the arts.