Reducing Head Pose Estimation Data Set Bias With Synthetic Data

Data set bias significantly impacts the performance of machine learning models, compromising their fairness, accuracy, and effectiveness. This issue is especially evident in the estimation of head pose, as current data sets suffer from a limited number of images, imbalanced data distributions, the high cost of annotation, and ethical concerns. Synthetic data offers a promising solution to address these challenges, but current semi-synthetic data sets fail to deliver satisfactory results, likely due to the limited realism of the generated faces and the heavily skewed pose distribution. In this paper, we examine the presence of data set biases in the most widely used head pose estimation benchmarks and create a data set of synthetic images using a generative model with explicit control over the head pose to mitigate this issue. Our experiments demonstrate that incorporating our synthetic images leads to improved generalization and accuracy.