violence-detection-yolov26

Violence Detection — YOLOv26-s

A single-class object detector that localizes violence in images/video frames. Fine-tuned from Ultralytics yolo26s.pt (COCO-pretrained).

Model details

Evaluation (validation set)

Training data

Trained on a custom YOLO-format dataset derived from surveillance/anomaly sources. The dataset-preparation scripts reference:

• ShanghaiTech Campus anomaly dataset
• UCF-Crime dataset

⚠️ These source datasets are released under their own academic/research-use terms and are not redistributed here. Only the fine-tuned weights are shared. Verify you have the right to use this model for your use case.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("image.jpg")          # or a video / webcam stream

for r in results:
    r.show()                           # display detections
    print(r.boxes.xyxy, r.boxes.conf)  # boxes + confidence

License

This model is derived from Ultralytics YOLO, which is AGPL-3.0. The weights are therefore distributed under AGPL-3.0. For closed-source/commercial use, obtain an Ultralytics Enterprise License.

Limitations

• Single-class detector — flags "violence" regions only; no fine-grained action labels.
• Trained on surveillance-style footage; may not generalize to other domains.
• Not validated for safety-critical or real-world deployment. Human review required.