I trained YOLOv4 model using darknet framework to detect weapons in images and videos.
I got the data from open images, I used 3 classes Handgun, Shotgun and Weapon, then I combined them together as single class "Weapon", So I had 1425 images for training and 285 images for validation.
I trained the model for 4000 epochs and got mAP about 65% with threshold=0.5
YOLOv4 Implemented in Tensorflow 2.0.
I Converted yolov4-custom.weights to format for tensorflow.
Download yolov4-custom.weights file: https://drive.google.com/file/d/1-bPteE8rV9ObeX4l28A5XWMQjWWtWsBG/view?usp=sharing
Example => https://youtu.be/Dcs_drMG9Rg