jjcao / spe2018

spe2018

Contents

• 1. Environment preparation
• 2. Data preparation
• 3. Training

1 Environment preparation

a) Install miniconda 3 [//]: # "[install miniconda3-4.5.4, which is for python 3.6, not 3.7.] (https://repo.continuum.io/miniconda/) you can try install latested miniconda3 with python 3.7" update it if necessary.

conda update conda

b) Prepare a python 3.6 environment Create a python 3.6 environment (denoted by py36), if the base environment of miniconda3 is not python 3.6, which is usually 3.7 now.

conda create -n py36 python=3.6

c) Intall some libraries: trimesh 2.33, pytorch >= 0.4, opencv. (trimesh installation, refer to https://trimsh.org/install.html)

    # 
    conda install -c anaconda numpy scipy

    # h5py for pytorch_HMR, matplotlib for pointnet, pyglet for trimesh    
    conda install -c conda-forge h5py matplotlib trimesh pyglet
 
    # pytorch 
    conda install pytorch torchvision -c pytorch

d) [It is not used now. so ignore this step.] Install pointnet++(pointnet2) for pytorch: https://github.com/erikwijmans/Pointnet2_PyTorch. It will compile some c++ code.

    mkdir build && cd build
    cmake .. && make
    git submodule update --init --recursive

    # tqdm and natsort for pointnet ++. 
    # conda install -c conda-forge tqdm natsort
    # pip install sacred

2 Data preparation

Refer to ./datageneration/readme.md

3 Training

a) in ./, unzip the model.zip Only neutral_smpl_with_cocoplus_reg.txt (112M) and neutral_smpl_mean_params.h5 are necessary.

b) Train

python ./train.py

mesh & pointcloud library or display

1. open3d 0.3 不可调试，显示极棒，只能读取某种ply（obj，和meshlab的ply都读不了）,改写费劲。
    conda install -c open3d-admin open3d
2. trimesh 2.33 不可调试，实用。
    python 2.7 ok, python 3.6 ok
    conda install -c conda-forge trimesh
    conda install -c conda-forge pyglet

3. mayavi
    its dependencies are unfortunately rather heavy and failed to install for python 3.6;
    works for python 2.7
    conda install -c anaconda mayavi

4. pymesh 没试，没时间，trimesh+mayavi暂时够了。