A PyTorch Reproduction of HCN
Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation. Chao Li, Qiaoyong Zhong, Di Xie, Shiliang Pu, IJCAI 2018.
- NTU RGB+D: Cross View (CV), Cross Subject (CS)
- SBU Kinect Interaction
- Action recognition
- Action detection
- Visdom supported.
Our code is based on Python3.5. There are a few dependencies to run the code in the following:
- Python >= 3.5
- PyTorch == 0.4.0
- Other version info about some Python packages can be found in
To transform raw NTU RGB+D data into numpy array (memmap format ) by this command:
python ./feeder/ntu_gendata.py --data_path <path for raw skeleton dataset> --out_folder <path for new dataset>
Not supported now.
Before you start the training, you have to launch visdom server.
python -m visdom
To train the model, you should note that:
--dataset_diris the parents path for all the datasets,
--numthe number of experiments trials (type: list).
python main.py --dataset_dir <parents path for all the datasets> --mode train --model_name HCN --dataset_name NTU-RGB-D-CV --num 01
To run a new trial with different parameters, you need to:
- Firstly, run the above training command with a new trial number, e.g,
--num 03, thus you will got an error.
- Secondly, copy a parameters file from the
./HCN/experiments/NTU-RGB-D-CV/HCN01/params.jsonto the path of your new trial
"./HCN/experiments/NTU-RGB-D-CV/HCN03/params.json"and modify it as you want.
- At last, run the above training command again, it will works.
python main.py --dataset_dir <parents path for all the datasets> --mode test --load True --model_name HCN --dataset_name NTU-RGB-D-CV --num 01
Load and Training
You also can load a half trained model, and start training it from a specific checkpoint by the following command:
python main.py --dataset_dir <parents path for all the datasets> --mode load_train --load True --model_name HCN --dataset_name NTU-RGB-D-CV --num 01 --load_model <path for trained model>
The expected Top-1 accuracy of the model for NTU-RGD+D are shown here (There is an accuracy gap. I am not the author of original HCN paper, the repo was reproduced according to the paper text and have not been tuned carefully):
Cross Subject (%)
Cross View (%)
- Loss & accuracy[CV]
 Chao Li, Qiaoyong Zhong, Di Xie, Shiliang Pu. Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation. IJCAI 2018.
 yysijie/st-gcn: referred for some code of dataset processing.