Train with 3D vertices only

[zgzxy001]

Hi,

Thank you for your great work! As my data only have 3D vertices (no joints), I want to use the vertice loss only. I modified the loss here to

loss = args.vertices_loss_weight * loss_vertices

But it raises the error:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn.

Do you have any idea about using vertice loss only? Thank you!

[WZS666]

Hello @zgzxy001 ,
May I know do you fix this problem.
And how you deal with the data without joints.
For me I'm confusing to generate the data by my own 3D mesh(.obj file) without joints.
Thanks !

[kevinlin311tw]

Hi all, sorry for the late reply.

In order to train with vertex loss only, there are some possible ways:
(1) Simplify use fake joint GTs in our training pipeline. Also, set "joint_loss_weight" to zero. Although it is hacky, this requires less modifications, and it might be easier to debug.
(2) Modify our network by removing any joint query tokens. Also, remove 2d/3d joint related loss functions.

In addition, you may need to tune a little bit for "vertices_loss_weight". The default setting may not work in this case.