The paper "Embodied Hands: Modeling and Capturing Hands and Bodies Together Supplementary Material" by Javier Romero, Dimitrios Tzionas, and Michael J. Black discusses the creation of the MANO hand model, which involves collecting and mirroring hand scans to create a balanced training dataset. The authors use a scanner specifically designed to capture hands with a fixed wrist position, ensuring accurate deformation data. By mirroring left-hand scans to appear as right-hand scans, they increase the size of the training dataset and remove handedness bias. This approach improves performance, as demonstrated in the experiments. The mirroring transformation is defined in the sagittal plane of the SMPL x coordinate system, and the transformation matrix \( M \) is used to mirror scan points, mesh templates, and shape blend shapes. Pose blend shapes, which depend on global coordinate frames, are modified to account for the mirror input transformation. The mirrored pose blend shapes are obtained by applying the rotation un-mirroring transformation to each 3x3 input block in the pose blend shape matrix. The evaluation section presents the generalization plot for the MANO hand model trained on different datasets. The plot shows that the model trained on an augmented dataset (right and mirrored left-hand scans) performs better than the model trained only on right-hand scans, indicating the effectiveness of the augmented dataset approach. The mean scan-to-mesh error for the full space is significantly reduced from 2.90 mm to 1.01 mm after training.The paper "Embodied Hands: Modeling and Capturing Hands and Bodies Together Supplementary Material" by Javier Romero, Dimitrios Tzionas, and Michael J. Black discusses the creation of the MANO hand model, which involves collecting and mirroring hand scans to create a balanced training dataset. The authors use a scanner specifically designed to capture hands with a fixed wrist position, ensuring accurate deformation data. By mirroring left-hand scans to appear as right-hand scans, they increase the size of the training dataset and remove handedness bias. This approach improves performance, as demonstrated in the experiments. The mirroring transformation is defined in the sagittal plane of the SMPL x coordinate system, and the transformation matrix \( M \) is used to mirror scan points, mesh templates, and shape blend shapes. Pose blend shapes, which depend on global coordinate frames, are modified to account for the mirror input transformation. The mirrored pose blend shapes are obtained by applying the rotation un-mirroring transformation to each 3x3 input block in the pose blend shape matrix. The evaluation section presents the generalization plot for the MANO hand model trained on different datasets. The plot shows that the model trained on an augmented dataset (right and mirrored left-hand scans) performs better than the model trained only on right-hand scans, indicating the effectiveness of the augmented dataset approach. The mean scan-to-mesh error for the full space is significantly reduced from 2.90 mm to 1.01 mm after training.