Towards End-to-End 3D Human Avatar Shape Reconstruction from 4D Data

We propose a novel approach for human avatar shape reconstruction from point cloud sequences. Our approach requires no pretrained human shape model and generates a subject specific deformable shape model. Contrary to existing template-based methods, we optimize all model parameters jointly on the GPU with an off-the-shelf optimizer to deal with a large amount of free variables. Preliminary qualitative evaluation shows promising results.