Unofficial-Implement-of-Openposeγγ
You can check the full result on YouTube or bilibiliγγ
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 3.6γγ
γTraining
-
Download vgg19 weights file here or ιΎζ₯: https://pan.baidu.com/s/1t6iouKeDZBZRRg4BXsv5GA ζεη : 4k1w and uzip to 'checkpoints/vgg/' (please create the path yourself).
-
Download COCO2017: 2017 Train images, 2017 Val images and 2017 Train/Val annotations here.
make sure have this structure:
-COCO/
γ-images/
γγ-train2017/
γγ-val2017/
γ-annotations/ -
Specify '--annot_path_train' and '--img_path_train' in train.py to your own 'COCO/annotations/' and 'COCO/images/'.
-
run train.py
python train.py
and install requirements follow the error and run again.
Test
Specify --checkpoint_path to the folder includes checkpoint files in run.py.γγ
- running on webcam
python run.py
γγ - running on video
python run.py --video images/video.avi
γγ - running on image
python run.py --image images/ski.jpg
γγ
pretrained model on COCO 2017 is available here or ιΎζ₯: https://pan.baidu.com/s/1jcwRsOuEaveZRBU50lP_cQ ζεη : mqkr, this checkpoint includes fine-tuned vgg weights.γγ