Our paper “G3FA: Geometry-guided GAN for Face Animation” was presented at the British Machine Vision Conference (BMVC) 2024 by Alireza Javanmardi from the Augmented Vision Department DFKI!
The British Machine Vision Conference (BMVC) is one of the major international conferences on computer vision and related areas and is organised by the British Machine Vision Association (BMVA). This year it was held in Glasgow from November 25th to 28th.
Our work presents a method for real-time face re-enactment from one single source image and addresses a key limitation of GAN-based face animation: geometric consistency.
G3FA enhances realism, especially in challenging head poses, by:
- Utilizing implicit 3D supervision through inverse rendering to extract depth and normal maps from 2D images, guiding the generator towards greater accuracy.
- Employing an ensemble of discriminators to incorporate the 3D properties, improving the generator’s understanding of human head structure.
- Leveraging face volume rendering with orthogonal ray sampling and volume rendering for high-quality image synthesis.
Check out the paper:
G3FA: Geometry-guided GAN for Face Animation, Alireza Javanmardi, Alain Pagani, Didier Stricker, British Machine Vision Conference (BMVC), 2024
Code is available at github.com/dfki-av/G3FA
Contact: Alireza Javanmardi, Alain Pagani