Comparative YOLO Training and Evaluation Outputs for Weapon Detection in Surveillance-Oriented Images

This dataset contains trained neural network weights, training configurations, and evaluation outputs for weapon detection in surveillance-oriented images using selected YOLO architectures. The work was created as a reproducible experimental output accompanying a bachelor’s thesis focused on the detection of handguns and knives using deep learning methods. The collection includes trained models based on YOLOv8, YOLOv10, YOLOv12, and YOLO26 in small, medium, and large variants. All models were trained under a unified experimental setup to enable direct comparison of their performance. The training procedure used a public annotated dataset containing two object classes: handgun and knife. The dataset includes:- trained model weights for all evaluated configurations,- training and execution scripts,- configuration files and parameter settings,- evaluation summaries and selected output metrics,- plots and visual reports generated during training and validation,- documentation supporting reproducibility of the experiments. The objective of this collection is to support reproducibility, benchmarking, and follow-up research in security-oriented computer vision, especially in tasks related to object detection in surveillance scenarios. The results show that medium-sized model variants achieved the most balanced performance under the given experimental conditions, while small and partially occluded objects remain challenging. This record does not republish the original third-party image dataset used for training unless explicitly permitted by its license. It provides derived research outputs only. Author contributions (CRediT): Lukáš Kopačka: Methodology, Software, Validation, Formal analysis, Writing – original draft. Kateřina Kolaříková: Data curation, Resources. Martin Haváček: Supervision, Project administration.