The combination of active learning and object detection is evaluated. On the object detection side, a detector from the well-know YOLO family, YOLOv7 is used. The active learning cycle is set as a pool-based algorithm, finding the most informative samples by uncertainty sampling. This framework is set up in the case of an environmental study, the detection of marine animals in aerial images. Moreover, the dataset has a hight foreground-background imbalance.
The results of this study shows that performance of object detection without active learning reaches at mAP0.5 of 61% against 55% with active learning. But the integration of active learning shows other advantages. Among it, needing less non-empty images to reach this score.
Keywords - Active Learning, Object Detection, YOLO, Machine Learning, environmental study