Active Object Detection with Epistemic Uncertainty and Hierarchical Information Aggregation

Despite the huge success of object detection, the training process still requires an immense amount of labeled data. Active learning has been proposed as a practical solution, but existing works on active object detection do not utilize the concept of epistemic uncertainty, which is an important metric for capturing the usefulness of the sample. Previous works also pay little attention to the relation between bounding boxes when computing the informativeness of an image. In this paper, we propose a new active object detection strategy that improves these two shortcomings of existing methods. We specifically consider a Bayesian framework and propose a new module termed model evidence head (MEH), to take advantage of epistemic uncertainty in object detection. We also propose hierarchical uncertainty aggregation (HUA), which realigns all bounding boxes into multiple levels and aggregates uncertainties in a bottom-up order, to compute the informativeness of an image. Experimental results show that our method outperforms existing state-of-the-art methods by a considerable margin.