cd lgss
python run.py config/xxx.py
run.py
is the main process, which is encapsulated and controlled by config
files.
Every running process creates a run folder containing experiments logs.
And we prepare a script gen_csv.py
to automaticly generate a csv file with desired setting and results records.
Please set up the config file in run
folders and set trainFlag
as False
and testFlag
as True
.
python run.py ../run/xxx/xxx.py
This will automatically load the best model in the experiment_name
folder.
If you wish to use a specific model, please use resume
to include the model path.
Modify the config file set dataset name as all
and change the mode
according to the preference
dataset = dict(
name = "all",
mode=['place','cast','act','aud'],
)
It is able to follow the following codes to process the data. Remember to read the argparase
to choose an ideal setting.
cd pre
python ShotDetect/shotdetect.py # Cut shot
python place/extrac_feat.py # Extract place feature
python audio/extrac_feat.py # Extract audio feature
And the full feature extraction is updated in movienet-tools
- pytube is to download YouTube video. Install with
pip install pytube3 --upgrade
- FFMPEG is to cut scene video and it is usually installed by your OS
cd pre
python demodownload.py ## Download a YouTube video with pytube
python ShotDetect/shotdetect.py --print_result --save_keyf --save_keyf_txt ## Cut shot
cd ../lgss
python run.py config/demo.py ## Cut scene
The video link in the pre/demodownload.py
might be invalid as time goes, and it may change to your own.
The demo code only use the image place feature for simplicity and casues inferior performance. It may change the threshold here to have a slight modification. The higher the threshold, the less scene it will generate. scene_dict, scene_list = pred2scene(cfg, threshold=0.8)
pre/ShotDetect
is developed based on PySceneDetect. The shot detector is optimized to suit for movie.
Parallel shot detection shotdetect_p.py
is also included for future usage.