This article describes a comparative assessment of four cutting-edge object detection architectures – YOLOv8, YOLOv12, Faster R-CNN, and SSD300 – for the application of automatic landmine detection. The goal is to evaluate detection accuracy, precision, recall, and inference speed in a low-data setting. The experiments were performed on the Landmine Detection Dataset of Roboflow, which was created by Northumbria University, and comprised around 1.2K labeled images. All the models were trained using annotations specific to its structure: YOLOv8 and YOLOv12 trained with YOLO-formatted annotations, whereas Faster R-CNN and SSD300 with COCO-formatted annotations. Those performance measures including mAP@0.5, Precision, Recall, F1-Score, and inference latency were examined. The findings show YOLOv8 had the greatest overall accuracy and velocity while YOLOv12 performed suboptimally even though it was newer in design – most probably because it relies on larger datasets for optimal feature learning. The research highlights dataset scale and annotation structure as being decisively important to detection performance and informing improved model choice for real-world landmine detection.