Improving Small Object Detection with Attention NMS

Abstract:
Intersection over Union (IoU) and No-maximum suppression (NMS) are the basic fundamental components of state-of-the-art anchor-based object detectors. However, they have some defects. IoU, as a positioning evaluation measure, is inconsistent with the offset loss function of the bounding box, which may lead to inaccurate object positioning. Therefore, we draw lessons from the idea of Hausdorff distance, propose the distance perception module, and combine it with the original IoU to form a new regression evaluation measure. NMS is the post-processing module of object detectors. However, in the case of high overlap, relatively small objects may be lost. Therefore, we propose the Attention NMS algorithm (A-NMS) by considering the context information of location confidence. Finally, we embed DIoU and A-NMS into the mainstream object detector respectively. On the benchmark data set MS-COCO, we find that the performance of these detectors has been significantly improved. Through experiments, we find that DIoU and A-NMS can complement each other, and their cooperative use can further improve the performance of the detector.