Readme for Tiny pose estimation

• Final update: 2018 Oct
• All right reserved @ Jaewook Kang 2018

About

The aim of this repository is to introduce a tiny pose estimation tutorial. This pose estimation model is based on single hourglass model.

• Alejandro Newell, Kaiyu Yang, Jia Deng, " Stacked Hourglass Networks for Human Pose Estimation," ECCV 2016. We implement the pose estimation model in Tensorflow.

Keywords

• Tensorflow
• Single hourglass model
• Human pose estimation
• Inverted bottleneck (Mobilenet v2)

Installation

Compiler/Interface Dependencies

• Tensorflow >=1.9
• Python2 <= 2.7.12
• Python3 <= 3.6.0
• opencv-python >= 3.4.2
• pycocotools == 2.0.0
• Cython == 0.28.4
• tensorpack == 0.8.0
• Tf plot == 0.2.0.dev0

Git Clone

git clone https://github.com/jwkanggist/tf-tiny-pose-estimation
# cd tf-tiny-pose-estimation/
git init

pip install -r requirement.txt
./sh_scripts/install_tensorflow_gpu.sh

Pycocotools Installation (Only for Win)

For OSX and Ubuntu, we can install pycocotool by pip

• Step 1) Download and install Microsoft Build Tools 2015
• Step 2) pip install cocoapi repository
• Step 3) Make pycocotools

git clone https://github.com/cocodataset/cocoapi
cd PythonAPI
make

How to Run

1. Downloading dataset

Downloading dataset from the AI challenger website and place the dataset on ./dataset.

• AI challenge dataset link

2. Training

python ./tf_modules/trainer.py

3. Monitoring by Tensorboard

tensorboard --logdir ./export/tf_logs

Components

./tfmodules/
├── dataset
│   └── ai_challenger
│
├── export
│   ├── train_setup_log
│   └── tf_logs
│
├── coco_dataload_modules
│   ├── testcodes
│   │   └── test_dataloader.py
│   ├── dataset_augment.py
│   └── dataset_prepare.py
│
├── data_loader.py
├── eval.py
├── model_builder.py
├── model_config.py
├── train_config.py
└── trainer.py

Log Data

• ./export/train_setup_log/: We log and store setup of each training run in this folder.
• ./export/tf_logs/run-yyyymmddHHmmss/train/: We log tensorboard summary of each training run in this folder.
• ./export/tf_logs/run-yyyymmddHHmmss/valid/: We log tensorboard summary of each validation run in this folder.
• ./export/tf_logs/run-yyyymmddHHmmss/pb_and_ckpt/: We save ckpt and pb files resulting from each training run.

Related Materials (All Korean)

• Jaewook Kang, Human pose estimation 101 with Tensorflow
• Jaewook Kang, Tensorflow Practical Project Configuration

Feedback

• Issues: report issues, bugs, and request new features
• Pull request
• Email: jwkang10@gmail.com

License

• Apach License 2.0

Authors information

• Jaewook Kang Ph.D.
• Personal website: link
• Facebook : link
• Linkedin : link