SlowFast-Network-pytorch

An easy PyTorch implement of SlowFast-Network "SlowFast Networks for Video Recognition".

We also complete a real-time action detection demo. The demo is orgnized as:

         Yolo v3  
            │    
            │
         deepsort  
            │      
            │ 
     SlowFast Network

Display the demo results

Run the demo on your own data

1.Clone the repository

git clone https://github.com/MagicChuyi/SlowFast-Network-pytorch.git

2.Download Yolo v3 model: https://pan.baidu.com/s/1tT2uzI44KD3zzAgMskU1Aw

3.Download DeepSort re-id model: https://pan.baidu.com/s/1D1_Lw_lq-O75xFX-zFEEbg

4.Download Pre-trained SlowFast Network model: https://pan.baidu.com/s/17GLB2k3VhPgRsVCadVmjaA

5.Modify the model path and your video path in video_demo.py.

6.Run video_demo.py.

Train your own model

1.Download AVA dataset.

2.Discard corrupted data.

3.Dataset should be orgnized as:

ava/ava
│    │ preproc_train  
│    │    │ clips
│    │    │ keyframes
│    │    │ ava_train_v2.2.csv
│    │ preproc_val  
     │    │ clips 
     │    │ keyframes 
     │    │ ...   

4.Modify the params in config.py and train_config.py.

5.Run train_video.py.

Requirements

python 3
PyTorch >= 1.0
tensorboardX
OpenCV

Code Reference:

[1] https://github.com/Guocode/SlowFast-Networks/
[2] https://github.com/potterhsu/easy-faster-rcnn.pytorch