The deep neural network has found widespread application in object detection due to its high accuracy. However, its performance typically depends on the availability of a substantial volume of accurately labeled data. Several active learning approaches have been proposed to reduce the labeling dependency based on the confidence of the detector. Nevertheless, these approaches tend to exhibit biases toward high-performing classes, resulting in datasets that do not adequately represent the testing data. In this study, we introduce a comprehensive framework for active learning that considers both the uncertainty and the robustness of the detector, ensuring superior performance across all classes. The robustness-based score for active learning is calculated using the consistency between an image and its augmented version. Additionally, we leverage pseudo-labeling to mitigate potential distribution drift and enhance model performance. To address the challenge of setting the pseudolabeling threshold, we introduce an adaptive threshold mechanism. This adaptability is crucial, as a fixed threshold can negatively impact performance, particularly for low-performing classes or during the initial stages of training. For our experiment, we employ the Southeast Area Monitoring and Assessment Program Dataset 2021 (SEAMAPD21), comprising 130 fish species classes with 28,328 image samples. The results show that our model outperforms the state-of-the-art method and significantly reduces the annotation cost. Furthermore, we benchmark our model's performance against a public dataset (PASCAL VOC07), showcasing its effectiveness in comparison to existing methods.