Summary: this paper targets to design a new effective active learning approach on object detection. Although active learning has been researched many years in classification, there is not many work in object detection. Besides, labels in object detection are often more hard to label than classification. To bridge this gap, authors design a new function to score the importance of images. Compared to previous work, they consider not only the similarity of selected samples but also a new complicated function to score. They evaluate their approach on two detection datasets and get larger improvement than other active learning approaches. The only point which make me confused is why they use xxx to score images (assumption).