I’m working on training the YOLO26m model to detect weapons, the model is very good at detecting weapons but it also detects dark spots or different objects as guns.
My dataset has approx. 10000 images, they are all well annotated with tight bounding boxes, I have also made sure that the weapon can be contextualized if only the bounding box is viewed.
My metrics: precision, recall, and mAP50 hot 80+, while mAP50-95 reaches to 50.
Im really stuck as to how do i make my model robust to not make these false detentions.
You would need to add the false positives to your dataset without labels and retrain