Decoupling Identity and Visual Quality for Image and Video Anonymization.

The widespread usage of computer vision applications in the public domain has opened the delicate question of image data privacy. In recent years, computer vision researchers have proposed technological solutions to anonymize image and video data so that computer vision systems can still be used without compromising data privacy. While promising, these methods come with a range of limitations, including low diversity of outputs, low-resolution generation quality, the appearance of artifacts when handling extreme poses, and non-smooth temporal consistency. In this work, we propose a novel network based on generative adversarial networks (GANs) for face anonymization in images and videos. The key insight of our approach is to decouple the problems of image generation and image blending. This allows us to reach significant improvements in image quality, diversity, and temporal consistency while making possible to train the network in different tasks and datasets. Furthermore, we show that our framework is able to anonymize faces containing extreme poses, a long-standing problem in the field.