Two examples:

1. 3D convolution neural network. (C3D)
2. Two stream network. (vgg16)

PS: tested through UCF-101 dataset.

• cson
• OpenCV (v3)
• tensorflow (v1.0)

Download pretrained model, e.g. vgg16 into ckpts directory.

mkdir ckpts
cd ckpts
aria2c 'http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz'
tar xvzf vgg_16_2016_08_28.tar.gz && rm vgg_16_2016_08_28.tar.gz

• You can use ./tools/_list_to_json.py to generate lists of videos with labels.

Tuning cson files in ./config directory to modify your model.

Call the multi_gpu_train.py script with json filename(without ".json").

e.g.

  python multi_gpu_train.py vgg16_rgb --clear

e.g.

  python multi_gpu_train.py vgg16_rgb

  python eval.py vgg16_rgb

  python -m tools.video_test vgg16_rgb