George Retsinas and colleagues propose SMIRK for precise reconstruction of 3D facial expressions from images. Addressing the limitation of expression diversity in training data, they replace differentiable rendering with a neural rendering module. This approach allows the reconstruction loss to focus on geometry and supplements training with synthetic images of varying expressions.

• Deficiency in expression diversity in training data impairs reconstruction accuracy.
• Neural rendering module introduced, emphasizing on face geometry optimization.
• Training data augmented wit synthetic images for varied expressions.
• Qualitative and quantitative assessments, along with perceptual evaluations, confirm SMIRK’s leading performance.

SMIRK’s ability to generate lifelike reconstructions is key for advancements in virtual reality, gaming, and telepresence. By improving the naturalness of virtual interactions, this technology could pave the way for more immersive and realistic user experiences. Read More