feature_extraction

Frame feature extraction

We provide generic code to extract the frame features.

ResNET

We used ResNET-152 pre-trained on ImageNet available here (place the .h5 file into this folder).

• Create and activate feature extraction environment:

conda env create -f src/feature_extraction/environment.yml

source activate Soccer-FeatureExtractor

• Extract frame features:

python src/feature_extraction main.py --ResNET

C3D

We used C3D pre-trained on Sports1M available here (place the .h5 file into this folder).

• Create and activate feature extraction environment:

conda env create -f src/feature_extraction/environment.yml

source activate Soccer-FeatureExtractor

• Extract frame features:

python src/feature_extraction main.py --C3D

I3D

• Create and activate feature extraction environment:

cd src/feature_extraction/i3d-feat-extract/

conda env create -f environment.yml

source activate i3d-feat-extract

• Get the original Kinetics-I3D:

Clone Kinetics-I3D:

git clone https://github.com/deepmind/kinetics-i3d.git

• update $PYTHONPATH:

export PYTHONPATH='<you_main_soccernet_github_path>/src/feature_extraction/i3d-feat-extract/kinetics-i3d':$PYTHONPATH

• Extract frame features:

python extract_i3d_spatial_features.py <you_main_soccernet_github_path>/data/ <you_main_soccernet_github_path>/data/

Due to the long computation time, it is recommended to parallelize the code for the I3D feature extraction. To do so, we used our cluster and the argument --jobid <jobid> (jobid from 0 to 499) to specify a single job per game. An example is provided with the file /src/feature_extraction/i3d-feat-extract/extract_I3D_football.sh using SLURM.