deep-high-resolution-net.TensorFlow

A TensorFlow implementation of HRNet-32.The dataset used to train the model is the AI Challenger dataset.

Just for fun! A 'famous' actor CXK in China and the keypoints estimated using the HRNet-32.

For more details, please refer to the paper and the dataset.

Environment

python 3.6 or higher
TensorFlow 1.11 or higher
PyCharm

How to Use

For Training

Download the AI Challenger dataset.
Convert the images in the AI Challenger dataset (train_images folder) to TFRecords by running the dataset.py. Please make sure that the dataset_root_path you used in the extract_people_from_dataset() function is the path of the AI Challenger dataset you saved in the previous step.
Run the train.py!

Please note that the structure of the HRNet is complicated. I trained the HRNet-32 network using 2 Nvidia Titan V graphics cards. As the limited of the graphics memory（16 GB）, the max batch size I used was 2, and it took around 30 hours to finish 1 epoch (189176 steps). The model files were uploaded to Google Drive and Baidu Cloud (Extraction code: 7hym).

For Testing

Finish the 4 steps in the training.
Make sure the dataset name, mode file name are corrected.
Run the test.py!

The result images will be saved in the _testimg folder. It will also generate the distances.npy and the classes.npy file, which will be used to calculate the AP50 and AP75 later.

For Evaluating

Run the evaluate.py.

It will print the AP50 and AP75 information in the command line.

For Debugging

If you encounter any problems, please try to run the temp.py file to see if it can work properly. It is a simple demo file that can predict the human pose in the cxk.mp4 file. Compare to other scripts, this one is easier to debug.