Question about coco_smplifyx_train.json about i2l-meshnet_release HOT 7 CLOSED

Yes actually, the code is doing that.
For the SMPL joint set (24 joints):

For the COCO joint set (17 joints):
np.dot(self.coco_joint_regressor, smpl_mesh_cam)

You can projects the 3D coordinates to 2D image space by

from i2l-meshnet_release.

Thank you very much for your reply!

from i2l-meshnet_release.

Hi, here is another question. I've converted the coco 3d mesh to 3d skeleton. But the style of the coco 3d keypoints is different from the annotation style of H36M. How can convert the coco style to the H36M style?
Thank you very much for your help!

from i2l-meshnet_release.

If you multiply data/Human36M/J_regressor_h36m_correct.npy to mesh, it produces H36M 3D joint coordinates.
If you multiply data/MSCOCO/J_regressor_coco_hip_smpl.npy to mesh, it produces MSCOCO 3D joint coordinates.

from i2l-meshnet_release.

Thank you very much for your help!
How did you obtain the smpl parameters of the MSCOCO train dataset? Could you please share the corresponding code?

from i2l-meshnet_release.

As I wrote in the paper and README, I used SMPLify-X (https://github.com/vchoutas/smplify-x).

from i2l-meshnet_release.

Okay! I'm very grateful for your help!

from i2l-meshnet_release.