Open-Pose-Only-Heatmap-Vectormap

Open Pose(Detecting Human Body Skeleton) network(Only heatmap and vectormap)

This project only could make vectormap and heatmap. So, I'll upload the project that could express the pose data applied the non-maxima suppression.

The Loss of this project is expressed here.

the left is heatmap loss and right is vectormap loss.

In my project, there are many python files.

1. For_parsing_mat_to_json.ipynb

this ipynb file makes .mat file to json file. When you download 'mpii_human_pose_dataset', there are annotation data. And the annotation data's format is .mat. So you convert this file format to json.

2. Image_resize.ipynb

this ipynb file makes image files to regular size(regular width x regular height). The reason why I make image files size regulary is for saving time. By making image file size regular, training time could be reduced. Because there are no time to resize image during training.

3. network.py and network.ipynb

this files are containing VGG-19 network.

4. Pose.ipynb

this ipynb file is for training. It contains class and util function to train.

Here is a output of this project. There are only vectormap and heatmap images.

input images

1. Heatmap images

2. Vectormap images