Hand posture parameter replacement

[Apeng-Rzp]

I'm sorry to bother you, but I met a problem when replacing the hand posture parameters. I notice that your hand attitude parameter is in the tensor form of [1*15*3*3]. I have the hand attitude parameter in the form of 1*15*3 tensor, converted by the angle_axis_to_rotation_matrix () ————https://github.com/kornia/kornia.

   def _compute_rotation_matrix(angle_axis, theta2, eps=1e-6):
        # We want to be careful to only evaluate the square root if the
        # norm of the angle_axis vector is greater than zero. Otherwise
        # we get a division by zero.
        k_one = 1.0
        theta = torch.sqrt(theta2)
        wxyz = angle_axis / (theta + eps)
        wx, wy, wz = torch.chunk(wxyz, 3, dim=1)
        cos_theta = torch.cos(theta)
        sin_theta = torch.sin(theta)

        r00 = cos_theta + wx * wx * (k_one - cos_theta)
        r10 = wz * sin_theta + wx * wy * (k_one - cos_theta)
        r20 = -wy * sin_theta + wx * wz * (k_one - cos_theta)
        r01 = wx * wy * (k_one - cos_theta) - wz * sin_theta
        r11 = cos_theta + wy * wy * (k_one - cos_theta)
        r21 = wx * sin_theta + wy * wz * (k_one - cos_theta)
        r02 = wy * sin_theta + wx * wz * (k_one - cos_theta)
        r12 = -wx * sin_theta + wy * wz * (k_one - cos_theta)
        r22 = cos_theta + wz * wz * (k_one - cos_theta)
        rotation_matrix = concatenate([r00, r01, r02, r10, r11, r12, r20, r21, r22], dim=1)
        return rotation_matrix.view(-1, 3, 3)

    def _compute_rotation_matrix_taylor(angle_axis):
        rx, ry, rz = torch.chunk(angle_axis, 3, dim=1)
        k_one = torch.ones_like(rx)
        rotation_matrix = concatenate([k_one, -rz, ry, rz, k_one, -rx, -ry, rx, k_one], dim=1)
        return rotation_matrix.view(-1, 3, 3)

    # stolen from ceres/rotation.h

    _angle_axis = torch.unsqueeze(angle_axis, dim=1)
    theta2 = torch.matmul(_angle_axis, _angle_axis.transpose(1, 2))
    theta2 = torch.squeeze(theta2, dim=1)

    # compute rotation matrices
    rotation_matrix_normal = _compute_rotation_matrix(angle_axis, theta2)
    rotation_matrix_taylor = _compute_rotation_matrix_taylor(angle_axis)

    # create mask to handle both cases
    eps = 1e-6
    mask = (theta2 > eps).view(-1, 1, 1).to(theta2.device)
    mask_pos = (mask).type_as(theta2)
    mask_neg = (~mask).type_as(theta2)

    # create output pose matrix
    rotation_matrix = eye_like(3, angle_axis, shared_memory=False)
    # fill output matrix with masked values
    rotation_matrix[..., :3, :3] = mask_pos * rotation_matrix_normal + mask_neg * rotation_matrix_taylor
    return rotation_matrix  # Nx3x3

The method adopted is to directly replace the hand parameters in the data with the converted hand parameters, but the final result is very bad. May I ask what should we pay attention to when replacing the hand posture parameters

pbar = tqdm(dataset)
    for data in pbar:
        print(data)
        data['left_hand_pose'] = left_hand_pose_apeng
        data['right_hand_pose'] = right_hand_pose_apeng
        .......

[YuliangXiu]

You can check rotation_converter.py for more rotation conversion operations.

[Apeng-Rzp]

sorry，the angle_axis_to_rotation_matrix () is not completely pasted and can be converted to and from your rotation_matrix_to_angle_axis ().

def angle_axis_to_rotation_matrix(angle_axis: Tensor) -> Tensor:
    r"""Convert 3d vector of axis-angle rotation to 3x3 rotation matrix.

    Args:
        angle_axis: tensor of 3d vector of axis-angle rotations in radians with shape :math:`(N, 3)`.

    Returns:
        tensor of rotation matrices of shape :math:`(N, 3, 3)`.

    Example:
        >>> input = tensor([[0., 0., 0.]])
        >>> angle_axis_to_rotation_matrix(input)
        tensor([[[1., 0., 0.],
                 [0., 1., 0.],
                 [0., 0., 1.]]])

        >>> input = tensor([[1.5708, 0., 0.]])
        >>> angle_axis_to_rotation_matrix(input)
        tensor([[[ 1.0000e+00,  0.0000e+00,  0.0000e+00],
                 [ 0.0000e+00, -3.6200e-06, -1.0000e+00],
                 [ 0.0000e+00,  1.0000e+00, -3.6200e-06]]])
    """
    if not isinstance(angle_axis, Tensor):
        raise TypeError(f"Input type is not a Tensor. Got {type(angle_axis)}")

    if not angle_axis.shape[-1] == 3:
        raise ValueError(f"Input size must be a (*, 3) tensor. Got {angle_axis.shape}")

    def _compute_rotation_matrix(angle_axis, theta2, eps=1e-6):
        # We want to be careful to only evaluate the square root if the
        # norm of the angle_axis vector is greater than zero. Otherwise
        # we get a division by zero.
        k_one = 1.0
        theta = torch.sqrt(theta2)
        wxyz = angle_axis / (theta + eps)
        wx, wy, wz = torch.chunk(wxyz, 3, dim=1)
        cos_theta = torch.cos(theta)
        sin_theta = torch.sin(theta)

        r00 = cos_theta + wx * wx * (k_one - cos_theta)
        r10 = wz * sin_theta + wx * wy * (k_one - cos_theta)
        r20 = -wy * sin_theta + wx * wz * (k_one - cos_theta)
        r01 = wx * wy * (k_one - cos_theta) - wz * sin_theta
        r11 = cos_theta + wy * wy * (k_one - cos_theta)
        r21 = wx * sin_theta + wy * wz * (k_one - cos_theta)
        r02 = wy * sin_theta + wx * wz * (k_one - cos_theta)
        r12 = -wx * sin_theta + wy * wz * (k_one - cos_theta)
        r22 = cos_theta + wz * wz * (k_one - cos_theta)
        rotation_matrix = concatenate([r00, r01, r02, r10, r11, r12, r20, r21, r22], dim=1)
        return rotation_matrix.view(-1, 3, 3)

    def _compute_rotation_matrix_taylor(angle_axis):
        rx, ry, rz = torch.chunk(angle_axis, 3, dim=1)
        k_one = torch.ones_like(rx)
        rotation_matrix = concatenate([k_one, -rz, ry, rz, k_one, -rx, -ry, rx, k_one], dim=1)
        return rotation_matrix.view(-1, 3, 3)

    # stolen from ceres/rotation.h

    _angle_axis = torch.unsqueeze(angle_axis, dim=1)
    theta2 = torch.matmul(_angle_axis, _angle_axis.transpose(1, 2))
    theta2 = torch.squeeze(theta2, dim=1)

    # compute rotation matrices
    rotation_matrix_normal = _compute_rotation_matrix(angle_axis, theta2)
    rotation_matrix_taylor = _compute_rotation_matrix_taylor(angle_axis)

    # create mask to handle both cases
    eps = 1e-6
    mask = (theta2 > eps).view(-1, 1, 1).to(theta2.device)
    mask_pos = (mask).type_as(theta2)
    mask_neg = (~mask).type_as(theta2)

    # create output pose matrix
    rotation_matrix = eye_like(3, angle_axis, shared_memory=False)
    # fill output matrix with masked values
    rotation_matrix[..., :3, :3] = mask_pos * rotation_matrix_normal + mask_neg * rotation_matrix_taylor
    print(rotation_matrix)
    return rotation_matrix  # Nx3x3

However, despite the correct conversion of posture parameters, [1*15*3]—[1*15*3*3]the full-obj is still very poor. I don't know why the face is particularly bad in my result