This code repository is a test repository for the paper "DenseGCN: A Multi-level and Multi-temporal Graph Convolutional Network for Action Recognition", used to test DenseGCN on the NTU RGB+D dataset and the NTU RGB+D 120 dataset.

Download Data

• Download the raw data of NTU-RGBD and NTU-RGBD120. Put NTU-RGBD data under the directory ./data/nturgbd_raw. Put NTU-RGBD120 data under the directory ./data/nturgbd120_raw.

Data Preparation

• For NTU-RGBD, preprocess data with python data_gen/ntu_gendata.py.
• For NTU-RGBD120, preprocess data with python data_gen/ntu120_gendata.py.

Update cuda extension

cd ./model/Temporal_shift
bash run.sh

model

The model weights file is saved in . /weights.

Test

If you wish to run the test program, run the following command.

• Testing the accuracy of DenseGCN in an x-view benchmark of the NTU RGB+D dataset.

  python test.py --mode ntu_xview
  

• Testing the accuracy of DenseGCN in an X-Sub benchmark of the NTU RGB+D dataset.

  python test.py --mode ntu_sub
  

• Testing the accuracy of DenseGCN in an X-Set120 benchmark of the NTU RGB+D 120 dataset.

  python test.py --mode ntu_xset120
  

• Testing the accuracy of DenseGCN in an X-Sub benchmark of the NTU RGB+D 120 dataset.

  python test.py --mode ntu_sub120