This project implement some two-stream convolutional network.

Origin Two-Stream
TSN
DTPP

UCF101 contains 101 actions, 13320 video clips.The dataset can be download hereUCF Dataset. About 6.93GB.|

ffmpeg can capture the video's image in one line. Opencv can also do this.

    'ffmpeg -i \"{}\" -vf scale=-1:240 \"{}/image_%05d.jpg\"'.format(video_file_path, dst_directory_path)

details can be find in video_jpg_ucf101_hmdb51.py

FlowNet2.0 is used here to get the flow. Use Docker to finish this part. I use the two job below.
NVIDIA-flownet2-python
lmb

every flo change into two img, u and v.

This part include the basebone model in the network.

Four models include here.

bninception
INceptionv4
ResNet
Inceptionv3

average_fusion and svm_fusion include here.

These module is based on pytorch.
Pretrained module is based on Cadene
This origin project is based on jerryhuang's project. two-stream-action-recognition