TAGPerson-Blender

This repo contains the rendering codes of the paper TAGPerson: A Target-Aware Generation pipeline for Person Re-identification.

We use BlenderPy to render images under parameterized control.

The training codes are in another project tagperson-reid, which uses PyTorch to train the models.

1. Get Started

1.1 Install Python Environment

Create the conda env:

conda create -n tagperson-blender python=3.7.10
conda activate tagperson-blender

Download the bpy wheel bpy-2.91a0-cp37-cp37m-manylinux2014_x86_64.whl, and then:

pip install bpy-2.91a0-cp37-cp37m-manylinux2014_x86_64.whl

Execute the post-processing command:

bpy_post_install

If encounter error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: libXrender.so.1: cannot open shared object file: No such file or directory

You should switch to root user and install the lib.

On Ubuntu:

sudo apt-get install libxrender1

On CentOS:

sudo yum install libXrender.x86_64

Note: install bpy by pip install bpy does not work for me.

Since we use plugins for blender to load models from MakeHuman, we need to put some plugin codes into the library path.

Put the plugin import_runtime_mhx2 and makewalk into the installation location of bpy, such as:

mv import_runtime_mhx2 makewalk ~/miniconda3/envs/tagperson-blender/lib/python3.7/site-packages/2.91/scripts/addons

The bpy-2.91a0-cp37-cp37m-manylinux2014_x86_64.whl, import_runtime_mhx2 and makewalk can be download from the link described in 1.2 Prepare Data

Install other python dependencies:

pip install -r requirements.txt

1.2 Prepare Data

In order to prepare the blender rendering, we need to prepare necessary data in data/ folder:

• data/makehuman/makehuman_2954_mhx2_standing02: 3D human models to render
• data/bg_collection: background images
• data/pose: pose structure json files
• data/attribute_distribution: The target-aware attribute_distribution information
• data/shield_v2: [Optional] The images used to create shield cases.

The attribute_distribution and shield_v2 are optional. All names of these paths can be modified, as long as they are correctly specified in the configuration file.

You can download these data from the link.

2. Demo

Suppose the 3D human models have been downloaded to data/makehuman folder.

We can run a demo script to render one person to a static image:

PYTHONPATH=. python demo/01_render_one_person.py --mesh_path data/makehuman/makehuman_2954_mhx2_standing02/193_Standing02.mhx2 --save_path tmp/sample_person.jpg

The rendered image is saved to tmp/sample_person.jpg.

You can also run the script with --demo_camera_azim option to observe the changes in rendering option values:

PYTHONPATH=. python demo/01_render_one_person.py --mesh_path data/makehuman/makehuman_2954_mhx2_standing02/193_Standing02.mhx2 --save_path tmp/sample_person.gif --demo_camera_azim

The rendered animation image is saved to tmp/sample_person.gif.

You can specify different human models by the --mesh_path option.

3. Render Multiple Images

Ensure that the data/makehuman/makehuman_2954_mhx2_standing02 folder exists, or create it by soft link:

ln -s data/makehuman/makehuman_2954_mhx2_standing02_mini data/makehuman/makehuman_2954_mhx2_standing02

To render a person ReID dataset, you can specific a config file and run:

PYTHONPATH=. python tools/batch_generate.py --config-file configs/TAGPerson/01_direct_transfer/tagp_base.yml

The images will be saved to the path specified in the OUTPUT_DIR_FOR_IMAGE config item.

If the 3D human models in the data/makehuman/ folder are not complete, you can use PROCESS.FIRST_INDEX and PROCESS.LAST_INDEX to limit the range:

PYTHONPATH=. python tools/batch_generate.py --config-file configs/TAGPerson/01_direct_transfer/tagp_base.yml PROCESS.FIRST_INDEX 193 PROCESS.LAST_INDEX 193

4. Rendering Options

We describe parts of the rendering options below.

We visualize some effects of changing different values for certain rendering options.

Rendering options for each experiment can be found in the configure files in this link.