This project is used to map 2D person joints to 3D SMPL model person joints. The input of SMPL model is pose and shape parameters. We want to generate joints by SMPL, and shape parameters don't affect joints' position, so we don't care shape parameters, just use pose parameters which size is 24*3.
We want to set up a mapping function, which can get input 2D joints , and return SMPL pose parameters, so that we can use these pose parameters to generate 3D joints. We use neural network to set up this mapping function.
We just use fully-connected layer. The model is similar with 3d-pose-baseline implemented by Pytorch.
We use SMPL model code from https://github.com/CalciferZh/SMPL , this code is much easier than official SMPL code.
SMPL/smpl_np.py
use numpy to create SMPL modeSMPL/smpl_tf.py
use tensorflow to create SMPL modecreat_sample.py
create train set and test set for our networkmodel.py
our networktrain.py
train networktest.py
test network
use MeshLab display 3D model
sudo apt install meshlab