2>pre_train sport1m where you can find https://www.dropbox.com/s/zvco2rfufryivqb/conv3d_deepnetA_sport1m_iter_1900000_TF.model?dl=0

3>you should have your own train and test list such as we share in annotation

4>you can use some sh in packege('list') to convert your video to frame with 5fps

5>use python train_pool5.py train your model

6>use python predict_c3d_fcn.py to get result

7>use MAP-IOU by jupyter notebook to get mAP