Being aware of localization accuracy by generating predicted-IoU-guided quality scores

Abstract

Localization quality estimation (LQE) methods benefit the post-process by additionally considering the prediction box’s localization accuracy. In this paper, we propose a more compatible detector called Classification-Localization-Quality (CLQ), which is not only applicable to general-distribution-based detection heads but also to delta-distribution-based heads. In this method, A lightweight and learnable LQE branch is designed to generate more accurate LQE scores. We also use an exponential factor and a weighted loss function respectively to optimize its effect and improve its training efficiency. To enhance task interaction during both the training and test phases, we merged this branch with the classification branch and trained them jointly. Experiment results show that CLQ achieves state-of-the-art performance at an accuracy of 47.8 AP and a speed of 11.5 fps with ResNeXt-101 as the backbone on COCO test-dev. We also extend our method to the ATSS baseline to evaluate the scalability. Codes are released at (https://github.com/PanffeeReal/CLQ).