Skip to content

Latest commit

 

History

History
26 lines (18 loc) · 4.14 KB

README.md

File metadata and controls

26 lines (18 loc) · 4.14 KB
Raw

Setup for data generation

Option 1: Automatic

We provide a bash script setup.sh that shows the steps to run to prepare the Blender environment and to download required assets. It is safer to run each line separately than running the whole script at once, because it is intended to provide a guide. You can try running it with bash setup.sh and get ready to Ctrl-C in case of issues. The steps are explained in comments inside the script as well as below in Option 2.

Option 2: Manual

Installation:

Download Blender. The provided code was tested with Blender2.92, with additional packages pip, scipy, joblib installed with these lines. Create an additional python3.7 environment to install a few packages openexr-python, ffmpeg with conda as in these lines. Then either copy the packages under Blender's bundled python or add those paths to environment variables PYTHONPATH, PATH and LD_LIBRARY_PATH as in these lines.

If you want to test our custom Blender package, you can skip the above steps and download it through here: blender-2.92.0-linux64_custom.tar.xz (244MB).

Note: At the time of running the experiments for the paper, and older version Blender2.78 was used and an additional python3.5 environment was created to install other dependencies such as scipy and openexr (due to known problem with pip in Blender2.78a). Python3.5 is no longer supported. To make a fresh version, the code has been updated to work with newer Blender versions. If you search the lines with # blender < 2.8x in the code, you will notice the modifications required to update from Blender2.78 to Blender2.92. The released datasets were created with the older version.

Assets:

Several asset files are required to run the synthetic data generation.

  1. Please follow the instructions at smpl_data/README.md to download these, make sure to comply with the associated licenses from SMPL and SURREAL.
  2. In addition, we run motion estimation methods to provide SMPL pose parameters to the data generation. We provide these estimated pose parameters within this project: vibe_ntu.tar.gz (3.9GB), hmmr_ntu.tar.gz (14GB), vibe_uestc.tar.gz (12GB), hmmr_uestc.tar.gz subset (420MB). Please cite the original datasets NTU and UESTC if you use these motion estimation data. See misc/motion_estimation to get an idea on how the motion estimates were obtained using VIBE or HMMR. Please comply with their licenses if you use code from VIBE or HMMR. Note that we have removed verts data to retain small storage, but the vertices can be obtained given the pose and shape parameters using the SMPL forward function.
  3. Finally, we used scripts in misc/background_crops to create background images from target datasets, you can download them here backgrounds_ntu.tar.gz (225MB) and backgrounds_uestc.tar.gz (323MB). You can also use LSUN images from SURREAL or any images to start.

Run data generation

Set the variables BLENDER_PATH and CODE_PATH in exe/run.sh, then run synthetic data generation on a single video clip:

bash exe/run.sh