Follow the latest coverage, related explainers and connected technology stories.
Through an experiment involving YOLO12, YOLO26, and RF-DETR, JetBrains shows that models pretrained on COCO may approach zero when transferred directly to specialized domains, while fine-tuning improves results to varying degrees. The experiment reveals that the similarity between the target domain and the training data, along with localization accuracy and inference time, matters more than relying on a model’s classification as the newest or best.
A re-evaluation of nine wheat head detection models shows that the overall mean mAP can conceal substantial performance differences between countries. China ranked first for all models, while areas of weakness differed between YOLO and RF-DETR, making cross-domain measurement essential before field deployment.