This repo is modified version for this code, which is source code for CVPR2018 paper.

Requirements

• PyTorch: 0.2 version (not test for other version)

pip3 install http://download.pytorch.org/whl/cu80/torch-0.2.0.post2-cp35-cp35m-manylinux1_x86_64.whl
pip3 install torchvision

• Python 3
• FFmpeg

wget http://johnvansickle.com/ffmpeg/releases/ffmpeg-release-64bit-static.tar.xz
tar xvf ffmpeg-release-64bit-static.tar.xz
cd ./ffmpeg-3.3.3-64bit-static/; sudo cp ffmpeg ffprobe /usr/local/bin;

UCF101

• Download videos and train/test splits from this site
• Use FFmpeg to extract frames from raw videos

python3 datasets/video_jpg_ucf101.py ucf101_video_dir ucf101_frames_dir

• Generate train/test video list

python3 datasets/ucf101_get_list.py ucf101_frames_dir splits/ucf101_train01_raw.txt
python3 datasets/ucf101_get_list.py ucf101_frames_dir splits/ucf101_test01_raw.txt

• finetune model from kinetics pretrain

mkdir logs; mkdir model;
bash ucf101_train.sh

• result

HMDB51

• Download videos and train/test splits from this site
• Use FFmpeg to extract frames from raw videos

python3 datasets/video_jpg_hmdb51.py hmdb51_video_dir hmdb51_frames_dir

• Generate train/test video list

python3 datasets/hmdb51_get_list.py hmdb51_frames_dir splits/hmdb51_train01_raw.txt
python3 datasets/hmdb51_get_list.py hmdb51_frames_dir splits/hmdb51_test01_raw.txt

• finetune model from kinetics pretrain

mkdir logs; mkdir model;
bash hmdb51_train.sh

• result
  • clip level: 53.45%
  • video leval: