• we choose yolov5 as an object detector instead of Faster R-CNN, it is faster and more convenient
• we use a tracker(deepsort) to allocate action labels to all objects(with same ids) in different frames
• our processing speed reached 24.2 FPS at 30 inference batch size (on a single RTX 2080Ti GPU)

Relevant infomation: FAIR/PytorchVideo; Ultralytics/Yolov5

1. create a new python environment (optional):

   conda create -n {your_env_name} python=3.7.11
   

2. install requiments:

   pip install -r requirements.txt
   

3. download weights file(ckpt.t7) from [deepsort] to this folder:

   ./deep_sort/deep_sort/deep/checkpoint/
   

4. test on your video:

   python yolo_slowfast.py --input {path to your video}
   

   The first time execute this command may take some times to download the yolov5 code and it's weights file from torch.hub, keep your network connection.

Thanks for these great works:

[1] Ultralytics/Yolov5

[2] ZQPei/deepsort

[3] FAIR/PytorchVideo

[4] AVA: A Video Dataset of Spatio-temporally Localized Atomic Visual Actions. paper

[5] SlowFast Networks for Video Recognition. paper

If you find our work useful, please cite as follow:

{   yolo_slowfast,
    author = {Wu Fan},
    title = { A realtime action detection frame work based on PytorchVideo},
    year = {2021},
    url = {\url{https://github.com/wufan-tb/yolo_slowfast}}
}