ahmedosman/supr

Official ECCV 2022 repository for SUPR: A Sparse Unified Part-Based Human Representation

supr's Issues

Foot Deformation Network Pose Input

Hey @ahmedosman et al., very excited for your work on SUPR, especially happy to see more effort in the lower body and feed region!
While replicating your results I ran into the issue of preparing the input for the provided foot deformation network, if I get it right the 13-joint pose parameter should be generated offline and passed as input along with the shape and translation right?
How did you infer this pose from in-the-wild images for evaluation?
Happy for any hints!

Broken hinge index list in constrained female/male model

Hey,

while testing the constrained body models I found a possible error in the kinematic tree.
While changing the value corresponding to the second hinge of the double hinge for the right middle finger, the first single hinge of the right middle finger is also affected.
It seems that one parameter is missing either for the double hinge or single hinge, I suspect that there's a problem in the kinematic tree that reuses the same parameter for both hinges.
The error doesn't occur in the constrained hand model.

Wrong axes in female/male constrained model

Hey,

In my work with the SUPR model, I have noticed that some of the constrainded axes do not seem to be at the correct joints.
This is most noticeable with the first joints of the ring fingers and the jaw.
After looking into the npy files, I simply swapped the constrained axes for the left and right hand as well as changing the first constrained axis for the jaw.
The fingers now seem to rotate correctly and the model can open the mouth.
The problem only occurs with the full body model and not with the separated parts.
I hope that I have noticed all the incorrect axes and that these will be fixed.

Betas

@ahmedosman
How can i defined betas or identify betas.

pose in constrained supr hand

When I use constrained hand model, I found hand mesh is not anatomical. I tried change the pose param of middle finger, but it is difficult to get a reasonable pose(espically when z axis value of mcp is ). It seems axes is not correct.
pose param: [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 1.5030, 0.0246, 1.0595, 1.6755, 1.3228, 0.0000, 0.9412, 2.0223, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]

I also have some quesetion about torch implementation:

1. https://github.com/ahmedosman/SUPR/blob/40e81d4b76eadb4c17de3cdfa2233c5cfe1b3654/supr/pytorch/utils.py#LL61 quat_feat() compute the normalized quat, why the res is v_cos-1.
2. torch_compute_rot_hinge is similar with formula, except the sign. If change the axes orientation, maybe this could be replaced by rodrigues.
3. quat2mat is Hamilton quat representation , while quat_feat is JPL. Why use two different representation in the code.

Hope for your relpy.

Bug in demo

Get this error when I run the demo:

UnboundLocalError: local variable 'v_shaped' referenced before assignment

This is in lib/python3.9/site-packages/supr/pytorch/supr.py

Left and right feet correspondence indices

Hi, thanks for open-sourced the model. Could you provide the indices for the left and right foot models?