Fight detection
A small model that looks at a video and tells if there is a fight in it.
▶ Try it online — upload a video and check it right in your browser.
Run it
git clone https://huggingface.co/Arystank01/fight-detection-x3d
cd fight-detection-x3d
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install --no-deps pytorchvideo==0.1.5
python app.py
Open http://127.0.0.1:7860, upload a video, press Check.
Or from the terminal:
python predict.py video.mp4
Works on CPU, uses GPU if you have one.
How it works
The model looks at 2 seconds of video at a time (16 frames) and gives a fight score from 0 to 1. It does this every 0.5 seconds. If 3 of the last 5 scores are above 0.6, it says there is a fight.
Training
- Model: X3D-M (about 3M parameters), pretrained on Kinetics-400
- Data: 3000+ short videos from CCTV and phones, about half with fights, half without
- Each training sample is a 2 second piece of a video, 16 frames, 224x224
- Best result after 3 epochs
To train it on your own videos, put them in folders like data/train/Fight/*.avi, data/train/NonFight/*.avi, data/val/Fight/*.avi, data/val/NonFight/*.avi and run:
python train.py data
Results
On 400 validation videos:
| Precision | Recall | F1 | |
|---|---|---|---|
| Fight | 88.8% | 87.5% | 88.2% |
| No fight | 87.7% | 89.0% | 88.3% |
Accuracy: 88.25%
| predicted no fight | predicted fight | |
|---|---|---|
| no fight | 178 | 22 |
| fight | 25 | 175 |
Limitations
- The model only sees the center square of the video. A fight at the edge of the frame can be missed.
- Crowds, hugging, sports and people moving fast close to each other can give false alarms.
- On your own cameras it can work worse than on the validation videos.
License
Free for personal, hobby and research use. Commercial use needs a paid license, contact me on GitHub (@arystan01). See LICENSE.
- Downloads last month
- 31