when 3D pose GT is taken as input, in Tab.4 the performance: 29.0 MPJPE and 23.0 PA-MPJPE on H36M. But in Tab.13, the performance: 13.9MPJPE 9.9 PA-MPJPE on H36M.
Why performance is inconsistent?

Hi, as written in the paper, MeshNet is trained with GT 3D pose in Table 13, while Pose2Mmesh is trained with GT 2D pose + synthesized errors in Table 4

I attempt to train MeshNet with GT pose as input on H36M (using SMPL skeleton).


In training set, we observe that MPVE: 6.3 mm, MPJPE: 11.9mm
But in testing set, MPVE: 20.9591mm, MPJPE: 33.7mm

Compared with using H36M skeleton (MPJPE: 13.9mm in Tab. 13), the results are very different. Can you tell me the code about MeshNet's individual training?

Train input: H36M skeleton
Test input: H36M skeleton
GT for evaluation: H36M skeleton

And also refer to yaml here.
I don't have the MeshNet training yaml file now, but I believe you can make it.

Thank you for your reply! Can you provide the pre-training model (only training MeshNet with 3D GT)?

Thank you for your reply! Can you provide the pre-training model (only training MeshNet with 3D GT)?

Hey I would like to ask do you train such pre-training model for MeshNet? I am also looking for that.

Hi all,

It's been a while after this project and I cannot find the pre-trained MeshNet model now. But it's easy to train MeshNet, referring to the code here: https://github.com/hongsukchoi/Pose2Mesh_RELEASE/blob/202ce98fce2bac74ff3787c7e30352a64d027950/lib/core/base.py

If you don't wanna train yourself, load the pre-trained Pose2Mesh and just use the MeshNet part by feeding the 3D pose to the MeshNet.