Rotation Augmentation for Head Pose Estimation Problem

The head pose estimation problem is one of the essential tasks of computer vision. Currently, many solutions use neural networks and show good results. Using data augmentations can significantly improve the quality of the trained neural network. Augmentations that affect the head rotation angles can increase the representativeness of head positions in the dataset. Horizontal flipping and rotating by a certain angle can achieve the desired effect. The rotation data augmentation is well known for the detection problem, but the only horizontal flip was used for the head pose estimation problem in previous papers. This paper proposes a modification of the image rotation augmentation for the head pose estimation problem, which corrects head rotation angles. Formulas for calculating the corrected rotation angles and bounding box are given in this paper. We conducted a series of computational experiments to test the effectiveness of the proposed augmentation. This paper compares proposed augmentation with different angles rotation and baseline, which applies only horizontal flip. The 300W - LP dataset was taken as a training dataset and the AFLW2000 dataset as a validation dataset. The pretrained Resnet50 was taken as a backbone. As the head branch for each angle, regression via classification with bin size equal to one was used. The best result is achieved by rotating at a random angle from -40 to 40 degrees. The accuracy of prediction by MAE metric angles was 3.85, which is close to the result of the best SOTA method with 3.83. The improvement in the accuracy of angle prediction relative to baseline is 22.1 % according to the MAE metric.