RankNet for accurate object detection with ranking bounding boxes

Abstract

This paper considers an architecture, referred to as RankNet, for improving object detection by addressing limitations in existing object detection methods as follows: (1) reweights the important samples considering the IoU and classification score of each positive and negative bounding box, (2) define a ranking score to measure the similarity between ground truth and predicted bounding box to indicate the ranks of bounding boxes for NMS procedure, and a metric called EIoU was proposed to calculate the ranking score. Extensive experiments on the standard COCO dataset show the effectiveness of the proposed method in multiple evaluation metrics, including multiple threshold AP metric. In particular, the proposed RankNet improves the AP of the bounding box 1.2, 2.3, 0.9 points compared to the Fast/Faster and Cascade R-CNN baseline, respectively.