XDVioDet

Official implementation of "Not only Look, but also Listen: Learning Multimodal Violence Detection under Weak Supervision" ECCV2020.

The project website is XD-Violence. The features can be downloaded from our project website.

where we oversample each video frame with the “5-crop” augment, “5-crop” means cropping images into the center and four corners. _0.npy is the center, _1~ _4.npy is the corners.

How to train

  • download or extract the features.
  • use make_list.py in the list folder to generate the training and test list.
  • change the parameters in option.py
  • run main.py

How to test

  • run infer.py

the model is in the ckpt folder.

Thanks for your attention!

GitHub

https://github.com/Roc-Ng/XDVioDet