Apple Inc.

Anna Volokitin

Brian Amberg

Stefan Brugger

Sebastian Martin

Ali Benlalah

Michael Tschannen

Advances in deep generative models (DGM) have led to the development of neural face video compression codecs that are capable of using an order of magnitude less data than 'traditional' engineered codecs. These “neural” codecs can reconstruct a target image by warping a source image to approximate the content of the target image and using a DGM to compensate for imperfections in the warped source image. The determined warping operation may be encoded and transmitted using less data (e.g., transmitting a small number of keypoints, rather than a dense flow field), leading to the bandwidth savings compared to traditional codecs. However, by relying on a single source image only, these methods can lead to inaccurate reconstructions. The techniques presented herein improve image reconstruction quality while maintaining bandwidth savings, via a combination of using multiple source images (i.e., containing multiple views of the first human subject) and novel feature aggregation techniques.

Growing community of inventors

Zurich, Switzerland

Anna Volokitin