Embodied Hands : Modeling and Capturing Hands and Bodies Together * * Supplementary Material * *

For the creation of the MANO hand model, we first collect a large number of scans of hands in isolation. These scans are obtained with a scanner configured specifically to capture hands with a fixed wrist position. This allows us to capture the nuances of hand deformation. After capturing this data for both right and left hands, we create a single augmented dataset by mirroring the left hand scans to appear as right ones. This approach increases the size of the training data and removes the bias introduced by the handedness of the subjects. In practical terms, it results in a performance improvement as shown in the experiments section. The augmented dataset enables us to train a single consistent hand model for both hands, i.e. we train the right hand model and generate the left one by mirroring. Model components which depend on the global coordinate frame, like the mesh template T̄, the shape blend shapes BS and the pose blend shapes BP , require mirroring. The rest of the components (e.g. the blend weightsW and joint regressor J ) remain untouched. We define the sagittal plane in SMPL, x , as our mirroring plane. This entails the following mirroring transformation