Skip to content

Latest commit

 

History

History
116 lines (92 loc) · 3.74 KB

CUSTOM_DATASET.md

File metadata and controls

116 lines (92 loc) · 3.74 KB

Guideline for preparing custom dataset

In this document, the process of preparing a custom dataset is elaborated. Any dataset that provides 3D mesh files (.obj) together with SMPL-X parameters can be used. We will describe the process using the THuman2.0 dataset as an example.

Dataset structure

$ROOT/datasets
└── THuman
   └── THuman2.0_Release
   └── THuman2.0_smplx   
   └── split_train.txt
   └── split_val.txt
   └── smplx_uv.obj  
   └── val 
      ├── img
      └── mask
      └── transform   
      └── position_map_uv_space
      └── position_map_uv_space_outer_shell_1
      └── position_map_uv_space_outer_shell_2
      └── position_map_uv_space_outer_shell_3
      └── position_map_uv_space_outer_shell_4
      └── visibility_map_uv_space
      └── visibility_map_uv_space_outer_shell_1
      └── visibility_map_uv_space_outer_shell_2
      └── visibility_map_uv_space_outer_shell_3
      └── visibility_map_uv_space_outer_shell_4

Render the dataset

Create the root directory to store your custom dataset

mkdir $ROOT/datasets/THuman

Download the raw dataset you want to process

First, download the target dataset and corresponding SMPL-X parameters.

  1. Create the directory to store the raw dataset
cd $ROOT/datasets/THuman
mkdir THuman2.0_Release
cd THuman2.0_Release

  1. Download the original THuman 2.0 dataset (THuman2.0_Release.zip) under $ROOT/datasets/THuman/THuman2.0_Release. Please complete the request form and send it to Yebin Liu ([email protected]) and cc Tao Yu ([email protected]) to request the download link.

  2. Unzip the dataset.

unzip THuman2.0_Release.zip
  1. Download the SMPL-X parameters (THuman2.0_smplx.tar.gz) corresponding to the original THuman 2.0 scans from here. Place the .tar.gz file under $ROOT/datasets/THuman and extract it.
tar -xvzf THuman2.0_smplx.tar.gz

  1. Create the dataset split files (.txt) and place them under $ROOT/datasets/THuman. You can refer to the dataset split files for GHG as a reference. The train split file can be downloaded here. The test split can be downloaded here.

  2. Download the smplx_uv.obj from the official SMPL-X website and place it under $ROOT/datasets/THuman.

7.Render RGB images and mask images.

python process_dataset/render_image.py

Generate adjusted SMPL-X obj

Please download the SMPL-X pkl files from the official SMPL-X website.

$ROOT/datasets
└── THuman
   └── models
      └── smplx
       ├── SMPLX_NEUTRAL.pkl
       ├── SMPLX_FEMALE.pkl
       └── SMPLX_MALE.pkl

During the rendering process, 3D scans are randomly transformed. To generate the matching SMPL-X obj, run the following command:

python process_dataset/generate_smplx_obj.py

Render position maps

Please install the NVDiffrast.

git clone https://github.com/NVlabs/nvdiffrast
pip install .

Render the position maps.

python process_dataset/render_position_map.py

Render visibility maps

Please make sure to modify the image_height and image_width to match the image plane shape of your dataset.

python process_dataset/render_visibility_map.py