/tf-dark-pose

This is an unofficial TensorFlow implementation of DARK Pose (Distribution Aware Coordinate Representation for Human Pose Estimation).

Primary LanguagePythonMIT LicenseMIT

TensorFlow implementation of DARK Pose

This is an unofficial TensorFlow implementation of DARK Pose (Distribution Aware Coordinate Representation for Human Pose Estimation).

It is based on the official PyTorch implementation ilovepose/DarkPose.

Requirements

  • Python 3.7
  • TensorFlow 2.1

Installation

  1. Create an anaconda environment.
conda create -n tf-dark-pose python=3.7 anaconda
  1. Activate the environment.
conda activate tf-dark-pose
  1. Install pip packages.
pip install -r requirements.txt

Training on COCO

  1. Download the images and annotation files (2017 train/val) from https://cocodataset.org/#download. Place the data in a directory structure as the following:
${COCO_ROOT}
├── images
|   ├── train2017
|   |   ├── 000000000009.jpg
|   │   ├── ...
|   │   
|   └── val2017
|       ├── 000000000139.jpg
|       ├── ...
|
└── annotations
    ├── person_keypoints_train2017.json
    └── person_keypoints_val2017.json
  1. Train the model.
python train.py configs/res50_128x96_d256x3_adam_lr1e-3.yaml --dataset_root ${COCO_ROOT}