Dynamic sample selection method based on classification prediction

Recent research has shown that sample selection in the training phase is crucial for improving the accuracy of object detection. For a long time, we always use manually designed IoU threshold to filter out low-quality candidates, but recently some work shows that dynamic sample selection is better for training an object detector. In our paper, we present a new adaptive dynamic sample selection strategy based on both IoU and Classification prediction. Our experiment applies in PASCAL VOC dataset, without any whistle, achieved 84% mAP. Also, our method doesn’t require any extra computation, which can guarantee high speed in the testing phase.