The paper introduces an online hard example mining (OHEM) algorithm for training region-based ConvNet detectors, addressing the challenge of imbalanced datasets in object detection. OHEM automatically selects hard examples, eliminating the need for heuristics and hyperparameters commonly used in region-based ConvNets. The method is applied to the Fast R-CNN framework, which is a popular and efficient object detector. Experiments on the PASCAL VOC 2007 and 2012 datasets show significant improvements in mean average precision (mAP), with OHEM achieving 78.9% and 76.3% mAP, respectively. The effectiveness of OHEM is further demonstrated on the more challenging MS COCO dataset, where it achieves state-of-the-art results. The paper also discusses the complementary benefits of combining OHEM with other recent advancements in object detection, such as multiscale testing and iterative bounding-box regression.The paper introduces an online hard example mining (OHEM) algorithm for training region-based ConvNet detectors, addressing the challenge of imbalanced datasets in object detection. OHEM automatically selects hard examples, eliminating the need for heuristics and hyperparameters commonly used in region-based ConvNets. The method is applied to the Fast R-CNN framework, which is a popular and efficient object detector. Experiments on the PASCAL VOC 2007 and 2012 datasets show significant improvements in mean average precision (mAP), with OHEM achieving 78.9% and 76.3% mAP, respectively. The effectiveness of OHEM is further demonstrated on the more challenging MS COCO dataset, where it achieves state-of-the-art results. The paper also discusses the complementary benefits of combining OHEM with other recent advancements in object detection, such as multiscale testing and iterative bounding-box regression.