Domain Analysis of End-to-end Recovery of Human Shape and Pose

This repository is part of the master thesis with the title: "Analysis of Domain Adaptability for 3D Human Pose Estimation with Explicit Body Shape Models" based on End-to-end Recovery of Human Shape and Pose.

This repository contains:

• reimplemented version of the original code
• a keypoint annotation tool to generate new keypoint regressors
  • integrates directly into the HMR framework
  • compatible with the SMPL model [2]

According to Kanazawa et al. [1], new keypoints can easily be incorporated with the mesh representation by specifying the corresponding vertexID. This feature makes the HMR framework very powerful. However, the definition of a general joint of the human body is more complex than just a single point on the surface. To address this problem, a new keypoint annotation tool has been developed that allows to generate a new keypoint regressor of a more complex keypoint. A detailed description of the annotation process and an installation guide can be found in keypoint_annotation_tool/README.md.

Requirements

• Python > 3.6
• Tensorflow > 2.1

Installation

Recommended Environment Setup

If you already have your environment set up, skip this and continue with Install Requirements.

1. Install virtualenv and virtualenvwrapper:

   pip3.x install --upgrade --user virtualenv
   pip3.x install --upgrade --user virtualenvwrapper
   

2. Add following to your .bashrc or .zshrc:

   export WORKON_HOME=$HOME/.virtualenvs
   # (optional) set export paths to your local python and virtualenv 
   # export VIRTUALENVWRAPPER_PYTHON=/usr/local/bin/python3.x 
   # export VIRTUALENVWRAPPER_VIRTUALENV=/usr/local/bin/virtualenv
   source /usr/local/bin/virtualenvwrapper.sh
   

3. set up the virtual environment

   mkvirtualenv hmr2.0
   workon hmr2.0
   pip install -U pip
   

Install Requirements

• check requirements.txt: choose between tensorflow (default) or tensorflow-gpu
• install requirements

  pip install -r requirements.txt
  

run demo

1. Create logs folder

   mkdir -p logs/paired(joints) && logs/unpaired
   

2. Download and unpack one of the pre trained models in the appropriate folder:

   • trained in paired setting (no SMPL parameter supervision due to missing dataset)

     • Base Model (LSP)
     • Toes Model (LSP + toes)
     • Total Capture Model (LSP) (original datasets plus Total Capture)
     • Final Model (LSP + toes) (original datasets plus Total Capture)
     • Tool Model (LSP) (original regressor with exchanged shoulder regressors )

   • trained in unpaired setting

     • Base Model (LSP)
     • Toes Model (LSP + toes)
     • Total Capture Model (LSP) (original datasets plus Total Capture)
     • Final Model (LSP + toes) (original datasets plus Total Capture)

3. (optional) All Models ending with (LSP + toes) need the toes regressors

   • link or copy the regressors folder from keypoint_annotation_tool to the models folder:

   cp -r /keypoint_annotation_tool/regressors
   # or
   cd models && ln -s ../keypoint_annotation_tool/regressors
   

4. Run demo cd src/visualise

   python demo.py --image=coco1.png --model=base_model --setting=paired\(joints\) --joint_type=cocoplus --init_toes=false
   

   

   python demo.py --image=lsp1.png --model=base_model --setting=paired\(joints\) --joint_type=cocoplus --init_toes=true
   

   

Note! No camera applied on the Mesh Overlay - Trimesh doesn't support orthographic projection)

Training

1. Convert datasets into TF Record format, see datasets_preprocessing/README.md
2. Update ROOT_DATA_DIR in src/main/config.py
3. (optional) Run src/visualise/notebooks/inspect_chekpoint.ipynb to update samples count in config for correct display of progress bar (requires jupyter installation)
4. Run training

   cd src/main
   python model.py > train.txt 2>&1 &!
   

Training takes up to 2 days on a RTX 2080 Ti GPU!

Evaluation

See eval/README.md

Source

[1]	Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. “End-to-end Recovery of Human Shape and Pose”. In: Computer Vision and Pattern Recognition (CVPR). 2018
[2]	Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. “SMPL: A Skinned Multi-Person Linear Model”. In: ACM Trans. Graphics (Proc. SIGGRAPH Asia) 34.6 (Oct. 2015), 248:1– 248:16.