Search

Accurate head pose estimation from 2D image data is an essential component of applications such as driver monitoring systems, virtual reality technology, and human-computer interaction. It enables a better determination of user engagement and attentiveness. The most accurate head pose estimators are based on Deep Neural Networks that are trained with the supervised approach and rely primarily on the accuracy of training data. The acquisition of real head pose data with a wide variation of yaw, pitch and roll is a challenging task. Publicly available head pose datasets have limitations with respect to size, resolution, annotation accuracy and diversity. In this work, a methodology is proposed to generate pixel-perfect synthetic 2D headshot images rendered from high-quality 3D synthetic facial models with accurate head pose annotations. A diverse range of variations in age, race, and gender are also provided. The resulting dataset includes more than 300k pairs of RGB images with corresponding head pose annotations. A wide range of variations in pose, illumination and background are included. The dataset is evaluated by training a state-of-the-art head pose estimation model and testing against the popular evaluation-dataset Biwi. The results show that training with purely synthetic data generated using the proposed methodology achieves close to state-of-the-art results on head pose estimation which are originally trained on real human facial datasets. As there is a domain gap between the synthetic images and real-world images in the feature space, initial experimental results fall short of the current state-of-the-art. To reduce the domain gap, a semi-supervised visual domain adaptation approach is proposed, which simultaneously trains with the labelled synthetic data and the unlabeled real data. When domain adaptation is applied, a significant improvement in model performance is achieved. Additionally, by applying a data fusion-based transfer learning approach, better results are achieved than previously published work on this topic.

This work explores the identity attribute of synthetic face samples derived from Generative Adversarial Networks. The goal is to determine if individual samples are unique in terms of identity, firstly with respect to the seed dataset that trains the GAN model and secondly with respect to other synthetic face samples. Two approaches are introduced to enable the comparative analysis of large sets of synthetic face samples. The first of these uses ROC curves to determine identity uniqueness using a number of large publicly available datasets of real facial samples to provide reference ROCs as a baseline. The second approach uses a thresholding technique utilizing again large publicly available datasets as a reference. For this approach, new metrics are introduced, and a technique is provided to remove the most connected data samples within a large synthetic dataset. The remaining synthetic samples can be considered as unique as data samples gathered from different real individuals. Several StyleGAN models are used to create the synthetic datasets, and variations in key model parameters are explored. It is concluded that the resulting synthetic data samples exhibit excellent uniqueness when compared with the original training dataset, but significantly less uniqueness when comparisons are made within the synthetic dataset. Nevertheless, it is possible to remove the most highly connected synthetic data samples. Thus, in some cases, up to 92% of the data samples in a 20k synthetic dataset can be shown to exhibit similar uniqueness to data samples taken from real public datasets.

Abstract: There is evidence that humans implicitly learn an average or prototype of previously studied faces, as the unseen face prototype is falsely recognized as having been learned (Solso & McCarthy, 1981). Here we investigated the extent and nature of face prototype formation where observers’ memory was tested after they studied synthetic faces defined purely in geometric terms in a multidimensional face space. We found a strong prototype effect: The basic results showed that the unseen prototype averaged from the studied faces was falsely identified as learned at a rate of 86.3%, whereas individual studied faces were identified correctly 66.3% of the time and the distractors were incorrectly identified as having been learned only 32.4% of the time. This prototype learning lasted at least 1week. Face prototype learning occurred even when the studied faces were further from the unseen prototype than the median variation in the population. Prototype memory formation was evident in addition to memory formation of studied face exemplars as demonstrated in our models. Additional studies showed that the prototype effect can be generalized across viewpoints, and head shape and internal features separately contribute to prototype formation. Thus, implicit face prototype extraction in a multidimensional space is a very general aspect of geometric face learning. [Copyright &y& Elsevier]