Cross-Domain Adaptation of YOLO-World for Open-Vocabulary Object Detection

We develop a unified multi-domain adaptation pipeline for adapting YOLO-World to heterogeneous medical and environmental imagery through annotation standardization, domain-specific fine-tuning, and multi-domain training. We systematically evaluate zero-shot, medical-only, and combined medical-environmental configurations, demonstrating substantial improvements in medical detection performance, with mAP@0.50 increasing from 0.60 to 0.89 and mAP@0.50:0.95 from 0.39 to 0.73. Furthermore, joint medical-environmental fine-tuning achieves 0.84 precision, 0.81 recall, and 0.87 mAP@0.50, demonstrating broader cross-domain generalization while retaining strong medical-domain performance. Finally, we provide a practical analysis of challenges such as small-object detection, occlusion, low contrast, prompt quality, and domain imbalance, while highlighting reproducibility considerations for future multi-domain open-vocabulary detection research.