HLA: Harmonized Label Assigner for Two-stage Oriented Object Detection

Abstract

The existing state-of-the-arts two-stage oriented object detectors have no significant improvement in the label assignment strategies, and the most widely-used one is the so-called Max IoU Assigner (MIA). In this paper, we first illustrate that MIA may cause matching conflicts in some cases, hinder the matching of ground-truth (GT) boxes with high-quality samples, which is extremely harmful to the training process. After that, we propose a Harmonized Label Assigner (HLA) for the oriented RPN, which can automatically harmonize the assignment priority of each GT box according to the corresponding number of candidate samples, solve the matching conflicts, and improve the detection accuracy of the two-stage oriented detectors. Finally, we implement the proposed HLA on Oriented R-CNN and conduct sufficient experiments on two public datasets (MAR20 and HRSC2016). Without tricks, our HLA significantly improves the detection accuracy of the detector to 83.97% mAP (on MAR20) and 90.42% mAP (on HRSC2016), respectively.