ZELF: A Zero-shot Ensemble Learning Framework for PPE detection in Industrial Workplaces

Object detection systems used in industrial environments typically rely on large-scale manually annotated datasets, which limit scalability and slow deployment. This paper proposes ZELF, a modular zero-shot ensemble learning framework that leverages multiple open-vocabulary object detectors to generate high-quality pseudo-labels without human annotation. ZELF aggregates predictions from YOLO-World, OWL-ViT, and Grounding DINO using class-aware Intersection-over-Union (IoU) based box merging and ensemble voting strategies. The resulting pseudo-labels are used to train a YOLOv8 student model. Experiments on the CHV PPE dataset show that ZELF achieves [email protected] scores of 0.966 for helmets, 0.605 for vests, and 0.437 for gloves when evaluated on pseudo-labeled data. When tested on ground-truth annotations, ZELF outperforms individual zero-shot detectors and achieves competitive performance, with [email protected] scores of 0.588 (helmet), 0.749 (vest), and 0.355 (glove). Compared to supervised baselines such as YOLOv7 and YOLOv8, ZELF demonstrates strong performance on visually distinct PPE categories while significantly reducing the need for manual labeling. These results highlight the effectiveness of ensemble-based zero-shot pseudo-labeling for scalable object detection in industrial settings.