hi, person_fit.ply has 6890 points but readme said it was stored in SMPL+H format about behave-dataset

[xiexh20]

Hi,

1. The SMPL model and SMPL-H model have the same topology, i.e. same number of vertices and faces. They only differ in the number of joint angle parameters.
2. Yes, same as SMPL mesh format.
3. I don't understand your problem. Please give more details. e.g. How do you read the ply file?

[JasonJ33JYS]

Thanks a lot for your quick reply! @xiexh20
For the question3, maybe my expression is not so good, I will try again.

I read the person_fit.ply in fit02 dir using plyfile.PlyData module. like below

ply_name = 'transfer_data\meshes\smplh\person_fit2.ply'
plydata = PlyData.read(ply_name)
print(plydata['vertex'].data.shape)
# (6890,)
print(plydata['face'].data.shape)
# (0,)

You can see that plydata has two elements: vertex and face, but face's shape is 0, which means there is no face in .ply file, and this is why: when I transform SMPL+H to SMPL using the repo below, the code got wrong, it tells me there is no face in .ply file.

I set face = [ ] in the smplx code, and it worked, but I don't know if the SMPL format result is right or not.

Thanks for your reply, again!

[JasonJ33JYS]

And other questions:

• Why there is only one trans in person_fit.pkl, we have four cameras (four views), should we have four trans?

• in k0.mocap.json, there are pose and shape estimated by FrankMocap, it is not so accurate compared to the real pose and shape. And the real pose and shape can be obtained by transforming person_fit.ply ( SMPL+H format) to person_fit.pkl ( SMPL format) using repo. Do I understand right?

• Why k0.mocap.json is stored in SMPL format, but person_fit.pkl in fit02 dir is stored in SMPL+H format?

• For the same static person, four views should have the same pose and shape? If it is, why four views have four kinds of pose and shape? k0.mocap.json k1.mocap.json k2.mocap.json k3.mocap.json

• How do you get person_fit.pkl?

• If I have SMPL pose and shape, and I put them in SMPL model to get 3D kp, and I project 3D kp to 2D kp, this 2D kp is more accurate or more not accurate than 2D kp openpose estimate.

I am new in 3d human pose estimation. So maybe my questions are too childish. Thanks for any help!
Best wishes! @xiexh20

[xiexh20]

Hi,

Sorry for the late reply. I am a bit busy these days.

1. The trans parameters is on the camera space of kinect 1. All four views are calibrated and it is known how to transform from one view to another. For example, here.
2. yes.
3. We use FrankMocap estimated pose to initialize our SMPL-H fitting. see Sec. 3 in the paper.
4. Again, k*.mocap.json is the result from FrankMocap. Each image is processed independently, the network prediction can have different outputs.
5. See paper sec.3 and supp. Sec2.3.
6. This I haven't tested. But the 3D keypoints are estimated from multi-view. So in theory it should be more accurate.