Separable 3D CNN with a spatio-temporal gating mechanism(S3D_G), proposaled in Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification(ECCV2018).
(Finish completely!!!)

This repo is to reimplement S3D_G, a powerful neural network for extracting spatial-temporal features from video/continuous frames. It produces very competitive result on several action classification benchmarks. For more detail, please access the link above.

• python3
• Pytorch1.0

$ git clone https://github.com/BIGJUN777/s3d_g_pytorch.git   

$ pip install -r requirements.txt

1. Download database: UCF101 or HMDB51.(Now just ucf101 supported!)
2. Create the folder and put the database into by making symlinks(or you can put databases into directly)

   $ mkdir dataset     		
   $ ln -s .../UCF101/UCF-101 dataset/UCF-101
   

3. Train the model. It will take some time to process the data in the first time running. Pay attention pls. Add -h to see more optional arguments.

   $ python train.py
   

4. Visualization:

   $ tensorboard --logdir=log
   

$ python inference.py --video path/to/a/video

• jfzhang95/pytorch-video-recognition
• rimchang/kinetics-i3d-Pytorch
• hassony2/kinetics_i3d_pytorch
• qijiezhao/s3d.pytorch