Confidence Estimation for Object Detection in Document Images

Deep neural networks are becoming increasingly powerful and large always require more labelled data to be trained. However, since annotating is time-consuming, it now necessary develop systems that show good performance while learning on a limited amount of data. These must correctly chosen obtain models still efficient. For this, the able determine which should annotated achieve best results. In this paper, we propose four estimators estimate confidence object detection predictions. The first two based Monte Carlo dropout, third one descriptive statistics last detector posterior probabilities. active framework, three significant improvement in for document physical pages text lines compared random selection images. We also proposed estimator can replace MC reducing computational cost without compromising performances.