ViSER: Video-Specific Surface Embeddings for Articulated 3D Shape Reconstruction