Follow the below link to install mmskeleton, mmdetection and openpose. (Must use "python setup.py develop --mmdet" to install the older version of mmdetection) https://github.com/open-mmlab/mmskeleton
pip install mmcv==0.4.3 -i https://pypi.tuna.tsinghua.edu.cn/simple conda install pytorch==1.3.0 torchvision cudatoolkit=10.1 -c pytorch
-
Rescale the training video to 340*256 and set frame rate to 30fps Useful ffmpeg comman under video folder:
for i in *.mp4; do ffmpeg -i "$i" -vf scale=340:256 "[path to output folder]/${i}"; done for i in *.mp4; do ffmpeg -i "$i" -filter:v fps=fps=30 "[path to output folder]/${i}"; done -
Put the training video to ./resource/data_example/
-
Modify the ./resource/data_example/category_annotation_example to add labels of different videos
-
run the command below in ./ to extract the skeleton data of the videos The extracted skeleton can be found in ./data/dataset_example/
mmskl configs/utils/build_dataset_example.yaml --gpus [num of gpus] -
Run the command below in ./ to train a new model.
mmskl configs/recognition/st_gcn/dataset_example/train.yaml -
Run the command below in ./ to train a new model.
mmskl configs/recognition/st_gcn/dataset_example/test.yaml -
Run the command below in ./ to get the visualized video.
a. If you want to get a visualized video, you can run the command below. The default path of the visualized video: "./visualized_video.avi" You the modify the file in "./mmskeleton/processor/demo_offline.py" in line 83.
python main.py demo_offline --openpose "[path to openpose build folder]" --video "[path to target video]"b. If you want to get a classification result of lists of video, you can run the command below after change variable visualization in "./mmskeleton/processor/demo_offline.py" in line 33 to False. The default path of the classification result: "./label_output.csv" You the modify the file in "./mmskeleton/processor/demo_offline.py" in line 43.
python main.py demo_offline --openpose "[path to openpose build folder]" --video "[path to target folder]"