IP-GAN: Learning Identity and Pose Disentanglement in Generative Adversarial Networks

PFA-GAN: Pose Face Augmentation Based on Generative Adversarial Network

“…Recently, generative adversarial network model learning (Tran et al, 2017;Tian et al, 2018;Cao et al, 2018a;Antoniou et al, 2018;Yin et al, 2017;Huang et al, 2017;Zeno et al, 2019a) demonstrated an outstanding ability to synthesize face images with new poses. Tran et al (2017) introduced Disentangled Representation Learning-Generative Adversarial Network (DR-GAN), where the model takes a face image of any pose as input and outputs a synthetic face, frontal or rotated with the target pose, even for extreme profiles (±90 • ).…”

“…Both TP-GAN and FF-GAN methods obtained impressive results on face frontalization, but they need explicit front-view annotations. Zeno et al (2019a) proposed a framework for Learning Identity and Pose Disentanglement in Generative Adversarial Networks (IP-GAN). To generate a face image of any specific identity with an arbitrary target pose, IP-GAN incorporates the pose information in the synthesis process.…”

“…The 2D-based methods rely on building a PCA model for a face shape to control only yaw rotations (Feng et al, 2017), while the 3D based methods synthesize face images with new variations of poses using a 3D morphable face model (Crispell et al, 2017;Blanz and Vetter, 1999;Zhu et al, 2016;Guo et al, 2017). In recent years, many learning-based methods have been proposed for face rotation, where most of them rely on a generative adversarial network (Tran et al, 2017;Tian et al, 2018;Cao et al, 2018a;Antoniou et al, 2018;Yin et al, 2017;Huang et al, 2017;Zeno et al, 2019a). For example, the methods DRGAN (Tran et al, 2017), CRGAN (Tian et al, 2018) and LB-GAN (Cao et al, 2018a) were proposed to rotate a face image around the yaw axis only.…”

“…The frontalization is a particular case of pose transformation often used to increase the accuracy of face recognition systems by rotating faces to the front view, such as in FF-GAN (Yin et al, 2017) and TP-GAN (Huang et al, 2017) works. Recently, Zeno et al (2019a) proposed IP-GAN framework to generate a face image of any specific identity with an arbitrary target pose by explicitly disentangling identity and pose representation from a face image.…”

“…indicating a head pose, 3DMM parameters). More specifically, Yin et al (2017), Huang et al (2017) need frontal face annotations, while (Tran et al, 2017;Tian et al, 2018;Cao et al, 2018a;Antoniou et al, 2018) need profile labels, while the IP-GAN (Zeno et al, 2019a) framework does not require any pose annotations. But despite this, it failed to learn disentangled representation of pose and identity on an unconstrained dataset of face images.…”