Hand_Keypoint_Detection

reference paper : http://openaccess.thecvf.com/content_cvpr_2017/papers/Simon_Hand_Keypoint_Detection_CVPR_2017_paper.pdf

1. The Network Architecture

2. Dataset To train this network, I use MPII Hand keypoint dataset. Datset Information is here. In annotation file, there is 21 keypoints information. And we could get 20 limb data by using 21 keypoints.

3. Training

The upper graph is loss of confidence map, and the lower graph is loss of PAFs.

4. Output I could get these output. First is image with representing keypoint.

And the second is image with representing PAFs with arrows.

The Confidencemap's size is [44x44x22]. And PAFs's size is [44x44x44]. So, I could get 22 confidencemaps and 44PAFs. Like this.

And, finally we could get these output

5. etc Information about this repository.

1. Demo_Hand.ipynb : If you trained this network, you could get PAFs and Confidencemaps by running session. Then, you use this ipynb file to estimate actual pose by using bipartite argorithm.

2. For_Demo_Hand.ipynb : This file is for Demo_Hand.ipynb.

3. Hand_Data_Processing.ipynb : You download MPII Hand keypoint dataset, and you could get raw annotation data not good for using. So, I made this file for using annotation data easily.

4. Training_Hand.ipynb : This file is for training.