Function get_transform_DART

motion_rep/retarget_motion.py:99–136

Compute the translation and rotation to align the SMPL-X model output to the canonical coordinate frame. Args: joints (torch.Tensor): SMPL-X joints of shape (seq_len, joints_num, 3). Returns: delta_transl (torch.Tensor): Translation vector of shape (3,). root_q

(joints)

    return smpl_data, joints

def get_transform_DART(joints):
    """
    Compute the translation and rotation to align the SMPL-X model output to the canonical coordinate frame.

    Args:
        joints (torch.Tensor): SMPL-X joints of shape (seq_len, joints_num, 3).

    Returns:
        delta_transl (torch.Tensor): Translation vector of shape (3,).
        root_quat_init (torch.Tensor): Rotation quaternion of shape (1, 4).
    """
    # Indices of the relevant joints (ensure these are correct for SMPL-X)
    pelvis_index = 0      # Pelvis (root joint)
    r_hip, l_hip, sdr_r, sdr_l = [2, 1, 17, 16]    # Right hip, Left hip, Right Shoulder, Left Shoulder

    device = joints.device

    # First frame joints positions
    joints_0 = joints[0]  # Shape: (joints_num, 3)

    # Step 1: Compute Rotation Quaternion

    # Compute x_axis (from left hip to right hip, projected onto xy-plane)
    x_axis = (joints_0[r_hip] - joints_0[l_hip])    # Shape: (3,)
    x_axis[2] = 0   # Project to the xy-plane (set z-component to zero)
    x_axis = x_axis / torch.norm(x_axis)  # Normalize

    # z_axis is pointing upwards (inverse gravity direction)
    z_axis = torch.tensor([0, 0, 1], dtype=torch.float32, device=device)

    # Compute y_axis as the cross product of z_axis and x_axis
    y_axis = torch.cross(z_axis, x_axis, dim=-1)
    y_axis = y_axis / torch.norm(y_axis)  # Normalize

    # Build rotation matrix R (from world frame to canonical frame)
    R_inv = torch.stack([x_axis, y_axis, z_axis], dim=1).T  # Shape: (3, 3)

    return R_inv

def apply_rotation(smpl_params, R):
    """