SMPL pose parameters

[melih-unsal]

Hi @akashsengupta1997 great congrats on your paper. When I have investigated your code I have found the pose parameter shape is (1,144) while standard SMPL has (1,72). Where are these extra 72 coming from? I need to render the SMPL with your parameters but the standard one waits for 72 for the pose. Can I get 72 pose parameters or are they irrelevant to each other?
Thanks.

[akashsengupta1997]

Hi sorry for the super late reply!

The standard SMPL model takes 72 pose parameters since they represent the rotations of 24 joints in axis-angle format i.e. each 3D rotation is represented by 3 numbers.

We instead use the 6 dimensional rotation representation proposed here: https://arxiv.org/abs/1812.07035, because it is a continuous representation of rotations (unlike axis-angle which is periodic and discontinuous) and makes training much easier. Thus there are 6 x 24 = 144 pose parameters.

6D rot representation can be converted to a usual 3x3 rotation matrix using:

def rot6d_to_rotmat(x):
    """Convert 6D rotation representation to 3x3 rotation matrix.
    Based on Zhou et al., "On the Continuity of Rotation Representations in Neural Networks", CVPR 2019
    Input:
        (B,6) Batch of 6-D rotation representations
    Output:
        (B,3,3) Batch of corresponding rotation matrices
    """
    x = x.view(-1,3,2)
    a1 = x[:, :, 0]
    a2 = x[:, :, 1]
    b1 = F.normalize(a1)  # Ensuring columns are unit vectors
    b2 = F.normalize(a2 - torch.einsum('bi,bi->b', b1, a2).unsqueeze(-1) * b1)  # Ensuring column 1 and column 2 are orthogonal
    b3 = torch.cross(b1, b2)
    return torch.stack((b1, b2, b3), dim=-1)

Then a 3x3 rotation matrix can be converted into axis-angle vector using cv2.Rodrigues().