In the Engineering department, we're using Deep Learning, including generative AI methods, to improve 3D animation production, focusing on retargeting, motion retrieval, motion style transfer, and text-based motion generation.

Our work involves 3D stylized characters, exploring different representation methods like skeleton definitions and 3D triangular meshes. Skeleton-based methods are good for retargeting but struggle with custom characters in production due to varying skeleton structures. The 3D mesh method, which uses faces and vertices, offers more flexibility but requires high memory and computational resources, limiting its use in dynamic and interactive scenarios.

The internship aims to review current mesh representation literature and develop a 3D mesh auto-encoder to efficiently capture object shapes and compress them into a latent representation. This could reduce computational demands, allowing for more complex motion models and interactions between multiple characters.