Home

Awesome

Unofficial-Implement-of-Openpose

<p align="left"> <img src="https://github.com/YangZeyu95/unofficial-implement-of-openpose/blob/master/readme/IMG_4063.GIF", width="720"> </p>　　

You can check the full result on YouTube or bilibili　　

An easy implement of openpose using TensorFlow.

Only basic python is used, so the code is easy to understand.

You can check the graph, internal outputs of every stage and histogram of every layer in tensorboard.

Original Repo(Caffe) : https://github.com/CMU-Perceptual-Computing-Lab/openpose.

The Dataloader and Post-processing code is from tf-pose-estimation.

Python 3.6　　

<p align="left"> <img src="https://github.com/YangZeyu95/unofficial-implement-of-openpose/blob/master/readme/graph_run%3D.png", width="720"> </p>　

Training

  1. Download vgg19 weights file here or 链接: https://pan.baidu.com/s/1t6iouKeDZBZRRg4BXsv5GA 提取码: 4k1w and uzip to 'checkpoints/vgg/' (please create the path yourself).

  2. Download COCO2017: 2017 Train images, 2017 Val images and 2017 Train/Val annotations here.
    make sure have this structure:
    -COCO/
    　-images/
    　　-train2017/
    　　-val2017/
    　-annotations/

  3. Specify '--annot_path_train' and '--img_path_train' in train.py to your own 'COCO/annotations/' and 'COCO/images/'.

  4. run train.py python train.py and install requirements follow the error and run again.

<p align="left"> <img src="https://github.com/YangZeyu95/unofficial-implement-of-openpose/blob/master/readme/loss2.svg", width="720"> </p>　　 　

Test

Specify --checkpoint_path to the folder includes checkpoint files in run.py.　　

pretrained model on COCO 2017 is available here or 链接: https://pan.baidu.com/s/1jcwRsOuEaveZRBU50lP_cQ 提取码: mqkr, this checkpoint includes fine-tuned vgg weights.