محمد ميرزائي

Active learning is recognized as one of the effective techniques for reducing the cost and time of labeling in the fields of computer vision and object detection. In active object detection, the selection of the initial labeled dataset is crucial, especially in complex datasets with high labeling costs and limited resources. This initial dataset holds even greater importance. This study presents a framework for active learning aimed at optimizing the initial training cycle for object detection tasks. Unlike traditional methods that select this dataset randomly, our approach leverages clustering based on features extracted from the YOLO-World model and refines the images using an "objectness" score obtained from the YOLO model. The proposed method focuses on identifying samples with diverse features and prioritizing training in areas where the foundation model exhibits the highest confidence. This helps the model better understand a variety of objects. To eva‎luate the effectiveness of the proposed method, two complex datasets, Aquarium and NWPU VHR-10, were utilized. These datasets challenge the foundation model, as it achieves low accuracy without training samples, and their diverse object sizes and challenging conditions make them suitable for testing our approach. In subsequent cycles, an inconsistency-based sampling technique is employed to maintain the stability of the training process. On the NWPU dataset, our clustering-based method improves the mean average precision (mAP50) by an average of 6.6% in image-based eva‎luations compared to random sampling. Furthermore, the clustering method combined with image refinement achieves an approximate 22% improvement in object-based eva‎luations. Experimental results demonstrate the advantages of using foundation models in the initial cycles of active learning. This achievement can play a significant role in advancing machine learning applications under resource and budget constraints and improving the efficiency of the learning process.

يادگيري فعال , مدل‌هاي بنيادين , تشخيص اشياء , يادگيري عميق، , برچسب‌گذاري داده‌ها

active learning , foundation models , object detection , deep learning , data labeling

Mohammad Mirzaee

Mohammad Reza Mohammadi