The script extracts frame level features from kinetics-i3d model (https://arxiv.org/abs/1705.07750). Dependencies: 1. OpenCV 2.4.13 compiled with GPU support 2. boost_python 1.65.0 3. Trained kinetics-i3d model (https://github.com/deepmind/kinetics-i3d) 4. TensorFlow 5. Sonnet Steps to Run: 1. Build CppInterface.so (CMakeList.txt is provided, please adjust paths, 'For ubuntu 16.04 run cmake -DCUDA_USE_STATIC_CUDA_RUNTIME=OFF') 2. run python ./extractor_lazy.py (it expects i3d.py (kinetics-i3d model code) in the same directory) 2.1 extractor.py expects list of video file names, 2.2 temporal_window for each frame default is 21 2.3 batch_size default is 1 2.4 dest_path (path to destination features) 2.5 clip_optical_flow_at default value is 20 2.6 base_path_to_chk_pts base path to model (kinetics-i3d) parameters A helper script FE_helper.py is also provided, please adjust it to your needs