Andreas Schneider

TL;DR: This study demonstrates the large potential of synthetic data for analyzing and reducing the negative effects of dataset bias on deep face recognition systems and shows that current neural network architectures cannot disentangle face pose and facial identity, which limits their generalization ability.

TL;DR: This work proposes a fully automated, probabilistic and occlusion-aware 3D morphable face model adaptation framework following an analysis-by-synthesis setup and proposes a RANSAC-based robust illumination estimation technique.

TL;DR: The proposed method allows valuable details about the generalization performance of different DCNN architectures to be observed and compared and reveals that the VGG-16 architecture outperforms the AlexNet architecture at face recognition tasks because it can much better generalize to unseen face poses, although it has significantly more parameters.

TL;DR: This work explores how synthetically generated data can be used to decrease the number of real-world images needed for training deep face recognition systems, and makes use of a 3D morphable face model for the generation of images with arbitrary amounts of facial identities and with full control over image variations.

TL;DR: In this paper, the authors empirically study the effect of different types of dataset biases on the generalization ability of deep convolutional neural networks (DCNNs) using synthetically generated face images, and reveal that the VGG-16 architecture outperforms the AlexNet architecture at face recognition tasks.