Jacobs-University/ICCV---AMASS---Archive-of-Motion-ICCV---AMASS---Archive-of-Motion-Capture-as-Surface-Shapes

Deep Learning Computer Vision models depend on massive datasets for their advance computations. There are multiple human motion capture datasets available which serves the purpose of doing the base computations but lack to be the ideal dataset in two major ways. Either, the datasets are very small and are constrained to only particular motion or the output generated by these datasets is not natural, that is, it’s not true to our humane movements. These problems are solved by AMASS (Archive of Motion Capture as Surface Shapes) which generates highly versatile SMPL (Multi Person Linear Model) model outputs which includes standard skeletal representation and a full surface body mesh. The MoSh algorithm which is used to generate the output is modified to incorporate the soft tissue dynamics and renamed to MoSh++. The dataset is then made richer by adding data from datasets like DMPL (Dynamic SMPL) and MANO (Hand Model with Articulated and Non-rigid Deformations). AMASS is able to generate output which is compatible with recent graphics engines, game engines and animations requirements.