Running the model on a video

[bilel-bj]

I would like to thank you for your interesting work. How can I run your model on a video to detect violence per frame? Thanks again for your interesting work.

[JimLee1996]

I've extracted the videos into frames using FFmpeg to save time on training,

To apply this model on videos, simply, load video using OpenCV or scikit-video. Next is to transform these frames into clips with shape 3x16x112x112 (please refer to transforms code and pytorch doc), then input them into model for inference.

[bilel-bj]

Thanks a lot for your kind reply. Could you please provide to the community the code for running the model on webcam. It will be very useful for the community. Thanks

[JimLee1996]

Please see https://github.com/JimLee1996/AVSS2019/tree/master/src/Demo

[bilel-bj]

Thanks a lot for providing the code to the community. I am interested in your code for research purposes. I found that you are the state of the art currently in violence detection in surveillance videos.

But there is one problem. I tested your pre-trained weight 'DenseNetLean_Kinetics.pth' on a video from the Hockey Fight dataset. Your model does not detect violence at all. It detects all the video as normal. Could you please provide the right weights?

[JimLee1996]

DenseNetLean_Kinetics.pth is pre-trained weight on kinetics and was not trained on violence datasets. I've upload the right weights trained on mix dataset to Demo.

[bilel-bj]

Thanks a lot for your great gift to the community!!!

[JimLee1996]

😉